Patent Application
20220345752
Video coding standards such as H.264/Advanced Video Coding (H.264/AVC), High-Efficiency Video Coding (HEVC) and Versatile Video Coding (VVC) share a similar (recursive) block-based hybrid prediction and/or transform framework. In such standards, to optimize the overall efficiency, individual coding tools like the intra/inter prediction, integer transforms, and context-adaptive entropy coding, are intensively handcrafted. These individual coding tools leverage spatiotemporal pixel neighborhoods for predictive signal construction, to obtain corresponding residuals for subsequent transform, quantization, and entropy coding. Neural networks on the other hand extract different levels of spatiotemporal stimuli by analyzing spatiotemporal information from the receptive field of neighboring pixels, essentially exploring highly nonlinearity and nonlocal spatiotemporal correlations. There is a need to explore improved compression quality using highly nonlinear and nonlocal spatiotemporal correlations.
Methods of lossy video compression often suffer from the compressed video having artefacts which severely degrade the Quality of Experience (QoE). The amount of distortion tolerated often depends on the application, but in general, the higher the compression ratio, the larger the distortion. Compression quality may be influenced by many factors. For example, the quantization parameter (QP) determines the quantization step size, and the larger the QP value, the larger the quantization step size, and the larger the distortion. To accommodate different requests of users, the video coding methods need the ability to compress videos with different compression qualities.
Although previous approaches involving deep neural networks (DNNs) have shown promising performance by enhancing video quality of the compressed video, it is a challenge for neural network-based (NN) quality enhancement methods to accommodate different QP settings. As an example, in previous approaches, each QP value is treated as an individual task and one NN model instance is trained and deployed for each QP value. In practice, different input channels have different QP values, e.g., chroma and luma components having different QP values. In such a situation, previous approaches require a combinatorial number of NN model instances. When more and different types off quality settings are added, the number of combinatorial NN models becomes prohibitively large. Moreover, a model instance trained for a specific setting of quality factors (QF) generally does not work well for other settings. While an entire video sequence usually has the same settings for some QF parameters, to achieve best enhancement effects, different frames may require different QF parameters. Therefore, methods, systems, and apparatuses that provide flexible quality control with arbitrary smooth settings of the QF parameters are required.
According to embodiments, a method for video enhancement based on neural network based loop filtering using meta learning is provided. The method may be executed by at least one processor, and the method may include receiving reconstructed video data; receiving one or more quality factors associated with the reconstructed video data; and generating enhanced video data with artefact reduction, based on the one or more quality factors and the reconstructed video data, using a neural network based loop filter, wherein the neural network based loop filter comprises neural network based loop filter parameters that include shared parameters and adaptive parameters.
According to embodiments, an apparatus for video enhancement based on neural network based loop filtering using meta learning is provided. The apparatus may include at least one memory configured to store program code; and at least one processor configured to read the program code and operate as instructed by the program code. The program code may include first receiving code configured to cause the at least one processor to receive reconstructed video data; second receiving code configured to cause the at least one processor to receive one or more quality factors associated with the reconstructed video data; and first generating code configured to cause the at least one processor to generate enhanced video data with artefact reduction, based on the one or more quality factors and the reconstructed video data, using a neural network based loop filter, wherein the neural network based loop filter comprises neural network based loop filter parameters that include shared parameters and adaptive parameters.
According to embodiments, a non-transitory computer-readable medium that stores instructions may be provided. The instructions, when executed by at least one processor, cause the at least one processor to receive reconstructed video data; receive one or more quality factors associated with the reconstructed video data; and generate enhanced video data with artefact reduction, based on the one or more quality factors and the reconstructed video data, using a neural network based loop filter, wherein the neural network based loop filter comprises neural network based loop filter parameters that include shared parameters and adaptive parameters.
Embodiments of the present disclosure describe methods, systems, and apparatuses for a quality-adaptive neural network-based loop filtering (QANNLF) for processing a decoded video to reduce one or more types on artefacts such as noises, blur, block effects, etc. In embodiments, a meta neural network-based loop filtering (meta-NNLF) method and/or process is described to adaptively compute quality-adaptive weight parameters of the underlying neural network-based loop filtering (NNLF) model based on based on the current decoded video and the QF of the decoded video, such as the Coding Tree Unit (CTU) partition, the QP, the deblocking filter boundary strength, the CU intra prediction mode, etc. According to embodiments of the present disclosure only one Meta-NNLF model instance may achieve effective artifact reduction over decoded videos with arbitrary smooth QF settings, including the seen settings in the training process and the unseen settings in actual application.
As shown in
The user device 110 may include one or more devices capable of receiving, generating, storing, processing, and/or providing information associated with platform 120. For example, the user device 110 may include a computing device (e.g., a desktop computer, a laptop computer, a tablet computer, a handheld computer, a smart speaker, a server, etc.), a mobile phone (e.g., a smart phone, a radiotelephone, etc.), a wearable device (e.g., a pair of smart glasses or a smart watch), or a similar device. In some implementations, the user device 110 may receive information from and/or transmit information to the platform 120.
The platform 120 may include one or more devices as described elsewhere herein. In some implementations, the platform 120 may include a cloud server or a group of cloud servers. In some implementations, the platform 120 may be designed to be modular such that software components may be swapped in or out. As such, the platform 120 may be easily and/or quickly reconfigured for different uses.
In some implementations, as shown, the platform 120 may be hosted in a cloud computing environment 122. Notably, while implementations described herein describe the platform 120 as being hosted in the cloud computing environment 122, in some implementations, the platform 120 may not be cloud-based (i.e., may be implemented outside of a cloud computing environment) or may be partially cloud-based.
The cloud computing environment 122 may include an environment that hosts the platform 120. The cloud computing environment 122 may provide computation, software, data access, storage, etc. services that do not require end-user (e.g., the user device 110) knowledge of a physical location and configuration of system(s) and/or device(s) that hosts the platform 120. As shown, the cloud computing environment 122 may include a group of computing resources 124 (referred to collectively as “computing resources 124” and individually as “computing resource 124”).
The computing resource 124 may include one or more personal computers, workstation computers, server devices, or other types of computation and/or communication devices. In some implementations, the computing resource 124 may host the platform 120. The cloud resources may include compute instances executing in the computing resource 124, storage devices provided in the computing resource 124, data transfer devices provided by the computing resource 124, etc. In some implementations, the computing resource 124 may communicate with other computing resources 124 via wired connections, wireless connections, or a combination of wired and wireless connections.
As further shown in
The application 124-1 may include one or more software applications that may be provided to or accessed by the user device 110 and/or the platform 120. The application 124-1 may eliminate a need to install and execute the software applications on the user device 110. For example, the application 124-1 may include software associated with the platform 120 and/or any other software capable of being provided via the cloud computing environment 122. In some implementations, one application 124-1 may send/receive information to/from one or more other applications 124-1, via the virtual machine 124-2.
The virtual machine 124-2 may include a software implementation of a machine (e.g., a computer) that executes programs like a physical machine. The virtual machine 124-2 may be either a system virtual machine or a process virtual machine, depending upon use and degree of correspondence to any real machine by the virtual machine 124-2. A system virtual machine may provide a complete system platform that supports execution of a complete operating system (“OS”). A process virtual machine may execute a single program, and may support a single process. In some implementations, the virtual machine 124-2 may execute on behalf of a user (e.g., the user device 110), and may manage infrastructure of the cloud computing environment 122, such as data management, synchronization, or long-duration data transfers.
The virtualized storage 124-3 may include one or more storage systems and/or one or more devices that use virtualization techniques within the storage systems or devices of the computing resource 124. In some implementations, within the context of a storage system, types of virtualizations may include block virtualization and file virtualization. Block virtualization may refer to abstraction (or separation) of logical storage from physical storage so that the storage system may be accessed without regard to physical storage or heterogeneous structure. The separation may permit administrators of the storage system flexibility in how the administrators manage storage for end users. File virtualization may eliminate dependencies between data accessed at a file level and a location where files are physically stored. This may enable optimization of storage use, server consolidation, and/or performance of non-disruptive file migrations.
The hypervisor 124-4 may provide hardware virtualization techniques that allow multiple operating systems (e.g., “guest operating systems”) to execute concurrently on a host computer, such as the computing resource 124. The hypervisor 124-4 may present a virtual operating platform to the guest operating systems, and may manage the execution of the guest operating systems. Multiple instances of a variety of operating systems may share virtualized hardware resources.
The network 130 may include one or more wired and/or wireless networks. For example, the network 130 may include a cellular network (e.g., a fifth generation (5G) network, a long-term evolution (LTE) network, a third generation (3G) network, a code division multiple access (CDMA) network, etc.), a public land mobile network (PLMN), a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a telephone network (e.g., the Public Switched Telephone Network (PSTN)), a private network, an ad hoc network, an intranet, the Internet, a fiber optic-based network, or the like, and/or a combination of these or other types of networks.
The number and arrangement of devices and networks shown in
A device 200 may correspond to the user device 110 and/or the platform 120. As shown in
The bus 210 may include a component that permits communication among the components of the device 200. The processor 220 may be implemented in hardware, firmware, or a combination of hardware and software. The processor 220 may be a central processing unit (CPU), a graphics processing unit (GPU), an accelerated processing unit (APU), a microprocessor, a microcontroller, a digital signal processor (DSP), a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), or another type of processing component. In some implementations, the processor 220 may include one or more processors capable of being programmed to perform a function. The memory 230 may include a random access memory (RAM), a read only memory (ROM), and/or another type of dynamic or static storage device (e.g., a flash memory, a magnetic memory, and/or an optical memory) that stores information and/or instructions for use by the processor 220.
The storage component 240 stores information and/or software related to the operation and use of the device 200. For example, the storage component 240 may include a hard disk (e.g., a magnetic disk, an optical disk, a magneto-optic disk, and/or a solid state disk), a compact disc (CD), a digital versatile disc (DVD), a floppy disk, a cartridge, a magnetic tape, and/or another type of non-transitory computer-readable medium, along with a corresponding drive.
The input component 250 may include a component that permits the device 200 to receive information, such as via user input (e.g., a touch screen display, a keyboard, a keypad, a mouse, a button, a switch, and/or a microphone). Additionally, or alternatively, the input component 250 may include a sensor for sensing information (e.g., a global positioning system (GPS) component, an accelerometer, a gyroscope, and/or an actuator). The output component 260 may include a component that provides output information from the device 200 (e.g., a display, a speaker, and/or one or more light-emitting diodes (LEDs)).
The communication interface 270 may include a transceiver-like component (e.g., a transceiver and/or a separate receiver and transmitter) that enables the device 200 to communicate with other devices, such as via a wired connection, a wireless connection, or a combination of wired and wireless connections. The communication interface 270 may permit the device 200 to receive information from another device and/or provide information to another device. For example, the communication interface 270 may include an Ethernet interface, an optical interface, a coaxial interface, an infrared interface, a radio frequency (RF) interface, a universal serial bus (USB) interface, a Wi-Fi interface, a cellular network interface, or the like.
The device 200 may perform one or more processes described herein. The device 200 may perform these processes in response to the processor 220 executing software instructions stored by a non-transitory computer-readable medium, such as the memory 230 and/or the storage component 240. A computer-readable medium may be defined herein as a non-transitory memory device. A memory device may include memory space within a single physical storage device or memory space spread across multiple physical storage devices.
Software instructions may be read into the memory 230 and/or the storage component 240 from another computer-readable medium or from another device via the communication interface 270. When executed, software instructions stored in the memory 230 and/or the storage component 240 may cause the processor 220 to perform one or more processes described herein. Additionally, or alternatively, hardwired circuitry may be used in place of or in combination with software instructions to perform one or more processes described herein. Thus, implementations described herein are not limited to any specific combination of hardware circuitry and software.
The number and arrangement of components shown in
Methods and apparatuses for video enhancement based on neural network based loop filtering using meta learning will now be described in detail.
This disclosure proposes a QANNLF mechanism to supports flexible quality control with arbitrary smooth settings of QF parameters for the compressed video. According to embodiments, meta-learning mechanism may be used to adaptively compute the quality-adaptive weight parameters of the underlying NNLF model based on the current decoded video and the QF parameters, enabling a single Meta-NNLF model instance to enhance decoded videos with different and arbitrary smooth QF settings.
Embodiments of the present disclosure relate to enhancing decoded videos to achieve effective artifact reduction over decoded videos with arbitrary smooth QF settings, including the seen settings in the training process and the unseen settings in actual application.
Generally, a video compression framework may be described as follows. Given an input video comprising of plurality of image inputs x1, . . . xT where each input image xt may be of size (h,w,c), may be an entire frame or a micro-block in an image frame such as a CTU where h, w, c are a height, a width, and a number of channels, respectively. Each image frame may be a color image (c=3), a gray-scale image (c=1), an rgb+depth image (c=4), etc. To encode video data, in a first motion estimation step, the input image(s) may be further partitioned into spatial blocks, each blocks partitioned into smaller blocks iteratively, and a set of motion vectors mt between a current input xi and a set of previous reconstructed inputs {}t-1 is computed for each block. The subscript t denotes the current t-th encoding cycle, which may not match the time stamp of the image input. Additionally, {
}t-1 may contain reconstructed inputs from multiple previous encoding cycles, such that the time difference between inputs in {
}t-1 may vary arbitrarily. Then, in a second motion compensation step, a predicted input {tilde over (x)}t may be obtained by copying the corresponding pixels of the previous {
}t-1 based on motion vectors mt. Then, a residual rt between the original input xt and the predicted input {tilde over (x)}t may be obtained. Then a quantization step may be performed where the residual rt may be quantized. According to embodiments, transformations such as DCT where the DCT coefficients of rt are quantized are performed prior quantizing the residual rt. A result of the quantization may be a quantized ŷt. Then both the motion vectors mt and quantized ŷt are encoded into bitstreams using entropy coding and sent to decoders. On the decoder side, the quantized ŷt may be dequantized to obtain the residual rt which is then added back to the predicted input {tilde over (x)}t to obtain reconstructed input {circumflex over (x)}t. Without limitations, any method or process may be used for dequantization, such as inverse transformations like IDCT with the dequantized coefficients. Additionally, without limitation, any video compression method or coding standard may be used.
In previous approaches, one or multiple enhancement modules may be selected to process reconstructed {circumflex over (x)}t, including Deblocking Filter (DF), Sample-Adaptive Offset (SAO), Adaptive Loop Filter (ALF), Cross-Component Adaptive Loop Filter (CCALF), etc, to enhance the visual quality of the reconstructed input {circumflex over (x)}t.
Embodiments of the present disclosure are directed to further improving the visual quality of the reconstructed input {circumflex over (x)}t. According to embodiments of the present disclosure, a QANNLF mechanism may be provided for enhancing the visual quality of the reconstructed input {circumflex over (x)}t of a video coding system. The target is to reduce artifacts such as noises, blur, blocky effects in {circumflex over (x)}t resulting in a high-quality {circumflex over (x)}th. More specifically, a meta-NNLF method may be used to compute {circumflex over (x)}th with only one model instance that may accommodate multiple and arbitrary smooth QF settings.
According to embodiments of the present disclosure, the proposed QANNLF and meta-NNLF may be used in combination with one or multiple of the previously mentioned additional components (i.e., DF, SAO, ALF, CCALF etc.) to improve the visual quality of the reconstructed input {circumflex over (x)}t. As an example, the reconstructed input {circumflex over (x)}t may go through DF first, followed by the Meta-NNLF, further followed by SAO and ALF. The order of how these components may be change arbitrarily. In one embodiment, the proposed method may be used alone as a replacement of all the other components to enhance the visual quality of reconstructed input {circumflex over (x)}t to obtain enhanced {circumflex over (x)}th.
Embodiments of the proposed meta-NNLF mechanism may operate on frame, tile, slice, CTU and/or CU level, in combination with one or multiple of the above mentioned additional components (e.g., DF, SAO, ALF, CCALF etc.). In other words, reconstructed input {circumflex over (x)}t may be one entire frame when fed into the Meta-NNLF, but in some embodiments, the reconstructed input {circumflex over (x)}t may also be a block when fed into the Meta-NNLF, e.g., a tile, slice, CTU or CU.
As shown in
As shown in
In this disclosure, model parameters of an underlying NNLF model are separated into 2 parts θs, θa denoting Shared NNLF Parameters (SNNLFP) and the Adaptive NNLF Parameters (ANNLFP), respectively.
In
In
The embodiment of
Let reconstructed input {circumflex over (x)}t of size (h,w,c,d) denote the input of the Meta-NNLF system, where h, w, c, d are the height, width, number of channels, and number of frames, respectively. Thus, a number of d−1 (d−1≥0) adjacent frames of {circumflex over (x)}t may be used together with {circumflex over (x)}t as input {circumflex over (x)}t to help generate the enhanced {circumflex over (x)}th. These multiple adjacent frames usually include a set of previous frames {′}, l<t, where each
′ may be decoded frame
or the enhanced frame {circumflex over (x)}th at a time l. Let Λt denote QF setting, each λl associated with each {circumflex over (x)}l′ to provide the corresponding QF information, and λt may be the QF setting for the current decoded frame {circumflex over (x)}t. The QF settings may include various types of quality control factors, such as the QP value, the CU intra prediction mode, the CTU partition, the deblocking filter boundary strength, the CU motion vector, and so on.
Let θs(i) and θa(i) denote the SNNLFP and ANNLFP for the i-th layer of the Meta-NNLF model 400, respectively. This is a general notation, since for a layer that may be completely shared, θa(i) is empty. For a layer that may be completely adaptive, θs(i) may be empty. In other words, this notation may be used for both embodiments of
An example embodiment of an inference workflow of the meta-NNLF model 400 for an i-th layer is provided.
Given the input {circumflex over (x)}t, and given the QF settings Λt, the Meta-NNLF method computes the enhanced {circumflex over (x)}th. Let f(i) and f(i+1) denote the input and output tensor of the i-th layer of the Meta-NNLF model 400. Based on a current input f(i) and θs(i), the SNNLFP Inference portion 412 computes a shared feature g(i) based on a shared inference function Gi(f(i), θs(i))) that may be modeled by a forward computation using the SEP in the i-th layer. Based on f(i), g(i), θa(i) and Λt, an ANNLFP Prediction portion 414 computes an estimated ANNLFP {circumflex over (θ)}a(i) for the i-th layer. The ANNLFP prediction portion 414 may be an NN, e.g., including convolution and fully connected layers, which predicts the updated {circumflex over (θ)}a(i) based on the original ANNLFP θa(i), the current input, and the QF settings Λt. In some embodiments, the current input f(i) may be used as an input to the ANNLFP prediction portion 414. In some other embodiments, the shared feature g(i) may be used instead of the current input f(i). In other embodiments, an SNNLFP loss may be computed based on the shared feature g(i), and a gradient of the loss may be used as input to the ANNLFP prediction portion 414. Based on the estimated ANNLFP {circumflex over (θ)}a(i) and the shared feature g(i), the ANNLFP inference portion 416 computes an output tensor f(i+1) based on an ANNLFP inference function Λi(g(i), {circumflex over (θ)}a(i)) that may be modeled by the forward computation using the estimated AEP in the i-th layer.
Note that the workflow described in
Assume there are a total of N layers for the meta-NNLF model 400, an output of a last layer may be the enhanced {circumflex over (x)}th.
Note that the meta-NNLF framework allows an arbitrary smooth QF settings for flexible quality control. In other words, the processing workflow described above will be able to enhance the quality of decoded frame with arbitrary smooth QF settings that may or may not be included in the training stage.
In a case, when the ANNLFP prediction portion 414 only performs prediction over a pre-defined set of QF settings with/without considering the input f(i), a meta-NNLF model reduces to a multi-QF NNLF model which uses one NNLF model instance to accommodate the enhancement of multiple pre-defined QF settings. Other reduced special cases may certainly be covered here.
As shown in
A training process aims at learning the SNNLFP θs(i) and ANNLFP θa(i), i=1, . . . , N for the Meta-NNLF model 400, as well as the ANNLFP Prediction NN (model parameters denoted as Φ).
In embodiments, a Model-Agnostic Meta-Learning (MAML) mechanism may be used for a training purpose.
For training, there may be a set of training data tr(Λi), i=1, . . . , K, where each
tr(Λi) corresponds to a training QF setting, and there are K training QF settings (thus K training data sets) in total. For training, there may be qqp different training QP values, qCTU different training CTU partitions, etc., and there may be a finite number of K=qqp×qCTU× . . . different training QF settings. Therefore, each training data set
tr(Λi) may be associated with each of these QF settings. In addition, there may be a set of validation data
val(Λj), j=1, . . . , P, where each
val(Λj) corresponds to a validation QF settings, and there are P validation QF settings in total. The validation QF settings may include different values from the training set. The validation QF settings may also have same values as those from the training set.
An overall training goal is to learn a meta-NNLF model so that it may be broadly applied to all (including training and future unseen) values of QF settings. The assumption being that an NNLF task with a QF setting may be drawn from a task distribution P(Λ). To achieve the training goal mentioned above, a loss for learning the meta-NNLF model may be minimized across all training data sets across all training QF settings.
The MAML training process may have an outer loop and an inner loop for gradient-based parameter updates. For each outer loop iteration, the task sampler 510 first samples a set of K′ training QF settings (K′≤K). Then for each sampled training QF setting Λi, the task sampler 510 samples a set of training data tr(Λi) from the set of training data
tr(Λi). Also, the task sampler 510 samples a set of P′ (P′≤P) validation QF settings, and for each sampled validation QF setting Λj, samples a set of validation data
val(Λj) from the set of validation data
val(Λj). Then for each sampled datum
∈
tr(Λi), a meta-NNLF forward computation may be conducted based on current parameters Θs, Θa, and Φ and the inner-loop loss generator 520 then computes an accumulated inner-loop loss
(θs, θa, Φ, Λi):
(θs,θa,Φ,Λi)=
({circumflex over (x)}t,θs,θa,Φ,Λi) (1).
The loss function L({circumflex over (x)}t, θs, θa, Φ, Λi) may include a distortion loss between a ground-truth image xtgt and the enhanced output {circumflex over (x)}th: D(xtgt, {circumflex over (x)}th) and some other regularization loss (e.g., auxiliary loss of distinguishing the intermediate network output targeting at different QF factors). Any distortion metric may be used, e.g., MSE, MAE, SSIM, etc., may be used as D(xtgt,{circumflex over (x)}th).
Then, based on the inner-loop loss (θs, θa, Φ, Λi), given step sizes αsi and αai as quality control parameters/hyperparameters for Λi, the inner-loop update portion 530 computes an updated task-specific parameter update:
{circumflex over (θ)}a=θa−Σi=1K′αai∇θ(θs,θa,Φ,Λi), (2); and
{circumflex over (θ)}s=θs−Σi=1K′αsi∇θ(θs,θa,Φ,Λi) (3).
Gradient ∇θ(θs, θa, Φ, Λi) and gradient ∇
(θs, θa, Φ, Λi) of the accumulated inner-loop loss
(θs, θa, Φ, Λi) may be used to compute an updated version of adaptive parameters {circumflex over (Θ)}a and {circumflex over (Θ)}s, respectively.
Then, a meta loss generator 540 computes an outer meta objective or loss over all sampled validation quality control parameters:
L(θs,θa,Φ)=Σj=1P′({circumflex over (θ)}s,{circumflex over (θ)}a,Φ,Λj), (4); and
({circumflex over (θ)}s,{circumflex over (θ)}a,Φ,Λj)=
L({circumflex over (x)}t,{circumflex over (θ)}s,{circumflex over (θ)}a,Φ,Λj) (5),
where L({circumflex over (x)}t, {circumflex over (θ)}s, {circumflex over (θ)}a, Φ, Λj) may be the loss computed for decoded frame {circumflex over (x)}t based on the meta-NNLF forward computation using parameters {circumflex over (θ)}s, {circumflex over (θ)}a, Φ, with QF setting Λj. Given step size βaj and βsj as hyperparameters for Ai, the meta update portion 550 updates the model parameters as:
θa=θa−Σj=1P′βaj∇θ({circumflex over (θ)}s,{circumflex over (θ)}a,Φ,Λj) (6); and
θs=θs−Σj=1P′βsj∇θ({circumflex over (θ)}s,{circumflex over (θ)}a,Φ,Λj) (7).
In some embodiments, Θs may not be updated in the inner loop, i.e., αsi=0, {circumflex over (Θ)}s=Θs. The non-updation helps to stabilize the training process.
As for parameters Φ of the ANNLFP Prediction NN, the weight update portion 560 updates them in a regular training manner. That is, according to the training and validation data tr(Λi), i=1, . . . , K′,
val (Λj), j=1, . . . , P′, based on the current θs, θa, Φ, we may compute loss L({circumflex over (x)}t, {circumflex over (θ)}s, {circumflex over (θ)}a, Φ, Λi) of all samples {circumflex over (x)}t∈
tr(Λi) and L({circumflex over (x)}t, {circumflex over (θ)}s, {circumflex over (θ)}a, Φ, Λj) for all samples {circumflex over (x)}t∈
val(Λj). And gradients of all these losses may be accumulated (e.g. added up) to perform parameter updates over Φ through regular back-propagation.
Embodiments of the present disclosure are not restricted to the above-mentioned optimization algorithm or loss functions for updating these model parameters. Any optimization algorithm or loss functions for updating these model parameters known in the art may be used.
For a case in which the ANNLFP prediction portion 414 of the Meta-NNLF model only performs prediction over the pre-defined set of training QF settings, the validation QF settings will be the same with the training ones. The same MAML training procedure may be used to train the above-mentioned reduced meta-NNLF model (i.e., a multi-QF-setting NNLF model that uses one model instance to accommodate compression effects of multiple pre-defined bitrates).
Embodiments of the present disclosure allows for using only one QANNLF model instance to accommodate multiple QF settings by using meta-learning. Additionally, embodiments of the present disclosure enable using only one instance of a meta-NNLF model to accommodate different types of inputs (e.g., frame level or block level, single image or multi-image, single channel or multi-channel) and different types of QF parameters (e.g., an arbitrary combination of QP values for different input channels, CTU partitions, the deblocking filter boundary strength, etc.)
As shown in
In some embodiments, the reconstructed video data may include a plurality of reconstructed input frames, and the methods described herein may be applied on a current frame of the plurality of reconstructed input frames. In some embodiments, the reconstructed input frames may be further broken down and used as the input to the meta-NNLF model.
Ins some embodiments, the one or more quality factors associated with the reconstructed video data may include at least one of a coding tree unit partition, a quantization parameter, a deblocking filter boundary strength, a coding unit motion vector, and a coding unit prediction mode.
In some embodiments, the reconstructed video data may be generated from a bitstream comprising decoded quantized video data and motion vector data. As an example, generating the reconstructed video data may include receiving a stream of video data including quantized video data and motion vector data. Then, generating the reconstructed video data may include dequantizing the stream of quantized data, using an inverse transformation, to obtain a recovered residual; and generating the reconstructed video data based on the recovered residual and the motion vector data.
The method 600A further may include generating enhanced video data with artefact reduction, based on the one or more quality factors and the reconstructed video data, using a neural network based loop filter, wherein the neural network based loop filter comprises neural network based loop filter parameters that include shared parameters and adaptive parameters. As an example, the meta-NNLF model 400 may be used, over a plurality of layers to generate enhanced video data with artefact reduction, based on the one or more quality factors and the reconstructed video data. In some embodiments, the generation of the enhanced video data using the neural network based loop filter may include operations 615-630. In some preferred embodiments, the generation of the enhanced video data using the neural network based loop filter may include operations 615-630 being repeated for a plurality of layers of the meta-NNLF model, example, meta-NNLF model 400.
Therefore, for each of a plurality of layers in the neural network based loop filtering using meta learning, at 615, the method may include generating shared features based on an output from a previous layer, using a first shared neural network loop filter having first shared parameters. As an example, SNNLFP inference portion 412 may generate shared features g(i) for an ith layer based on an output from the i−1 layer, using a first shared neural network loop filter having first shared parameters SNNLFP θse(i)
At 620, the method may include computing estimated adaptive parameters, based on the output from the previous layer, the shared features, first adaptive parameters from a first adaptive neural network loop filter, and the one or more quality factors, using a prediction neural network. As an example, the ANNLFP prediction portion 414 may compute estimated adaptive parameters {circumflex over (θ)}ae(i), based on the output from the i−1 layer, the shared features g(i), first adaptive parameters from a first adaptive neural network loop filter θae(i), and the one or more quality factors Λ, using a prediction neural network.
In some embodiments, the training of the prediction network may include generating a first loss for training data corresponding to the one or more quality factors, and a second loss for validation data corresponding to the one or more quality factors, based on the one or more quality factors, the first shared parameters, the first adaptive parameters, and prediction parameters of the prediction neural network, followed by updating the prediction parameters, based on gradients of the generated first loss and the generated second loss.
At 625, the method may include generating an output for a current layer, based on the shared features and the estimated adaptive parameters. As an example, the ANNLFP inference portion 416 may generate an output f(i+1) based on the shared features g(i) and the estimated adaptive parameters {circumflex over (θ)}ae(i).
At 630, the method may include generating enhanced video data, based on an output of a last layer of the neural network based loop filter.
In some embodiments, the method 600A may include post-enhancement processing including applying at least one of a deblocking filter, an adaptive loop filter, a sample adaptive offset, and a cross-component adaptive loop filter to the enhanced video data. As an example, post-enhancement processing may be performed subsequent to operation 630.
In some embodiments, the method 600B may include pre-enhancement processing including applying at least one of a deblocking filter, an adaptive loop filter, a sample adaptive offset, and a cross-component adaptive loop filter to the enhanced video data prior to generating the enhanced video data. As an example, pre-enhancement processing may be performed prior to operation 620.
As shown in
In some embodiments, the reconstructed video data may include a plurality of reconstructed input frames, and the methods described herein may be applied on a current frame of the plurality of reconstructed input frames. In some embodiments, the reconstructed input frames may be further broken down and used as the input to the meta-NNLF model.
Ins some embodiments, the one or more quality factors associated with the reconstructed video data may include at least one of a coding tree unit partition, a quantization parameter, a deblocking filter boundary strength, a coding unit motion vector, and a coding unit prediction mode.
In some embodiments, the reconstructed video data may be generated from a bitstream comprising decoded quantized video data and motion vector data. As an example, generating the reconstructed video data may include receiving a stream of video data including quantized video data and motion vector data. Then, generating the reconstructed video data may include dequantizing the stream of quantized data, using an inverse transformation, to obtain a recovered residual; and generating the reconstructed video data based on the recovered residual and the motion vector data.
The method 600B may further include generating enhanced video data with artefact reduction, based on the one or more quality factors and the reconstructed video data, using a neural network based loop filter, wherein the neural network based loop filter comprises neural network based loop filter parameters that include shared parameters and adaptive parameters. As an example, the meta-NNLF model 400 may be used, over a plurality of layers to generate enhanced video data with artefact reduction, based on the one or more quality factors and the reconstructed video data. In some embodiments, the generation of the enhanced video data using the neural network based loop filter may include operations 655-670. In some preferred embodiments, the generation of the enhanced video data using the neural network based loop filter may include operations 655-670 being repeated for a plurality of layers of the meta-NNLF model, example, meta-NNLF model 400.
At 655, the method may include generating a first shared output based on a previous combined output from a previous layer, using a first shared neural network loop filter having first shared parameters. At 660, the method may include generating a first adaptive output based on the previous combined output from the previous layer, using a first adaptive neural network loop filter having first adaptive parameters. At 665, the method may include combining the first shared output and the first adaptive output to generate a first combined output. At 670, the method may further include generating the enhanced video data, based on a combined output of a last layer of the neural network based loop filter.
The proposed methods may be used separately or combined in any order. Further, each of the methods (or embodiments), encoder, and decoder may be implemented by processing circuitry (e.g., one or more processors or one or more integrated circuits). In one example, the one or more processors execute a program that is stored in a non-transitory computer-readable medium.
In some implementations, one or more process blocks of
The foregoing disclosure provides illustration and description, but is not intended to be exhaustive or to limit the implementations to the precise form disclosed. Modifications and variations are possible in light of the above disclosure or may be acquired from practice of the implementations.
As used herein, the term component is intended to be broadly construed as hardware, firmware, or a combination of hardware and software.
It will be apparent that systems and/or methods, described herein, may be implemented in different forms of hardware, firmware, or a combination of hardware and software. The actual specialized control hardware or software code used to implement these systems and/or methods is not limiting of the implementations. Thus, the operation and behavior of the systems and/or methods were described herein without reference to specific software code—it being understood that software and hardware may be designed to implement the systems and/or methods based on the description herein.
Even though combinations of features are recited in the claims and/or disclosed in the specification, these combinations are not intended to limit the disclosure of possible implementations. In fact, many of these features may be combined in ways not specifically recited in the claims and/or disclosed in the specification. Although each dependent claim listed below may directly depend on only one claim, the disclosure of possible implementations may include each dependent claim in combination with every other claim in the claim set.
No element, act, or instruction used herein may be construed as critical or essential unless explicitly described as such. Also, as used herein, the articles “a” and “an” are intended to include one or more items, and may be used interchangeably with “one or more.” Furthermore, as used herein, the term “set” is intended to include one or more items (e.g., related items, unrelated items, a combination of related and unrelated items, etc.), and may be used interchangeably with “one or more.” Where only one item is intended, the term “one” or similar language is used. Also, as used herein, the terms “has,” “have,” “having,” or the like are intended to be open-ended terms. Further, the phrase “based on” is intended to mean “based, at least in part, on” unless explicitly stated otherwise.
This application is based on and claims priority to U.S. Provisional Patent Application No. 63/176,787, filed on Apr. 19, 2021, the disclosure of which is incorporated by reference herein in its entirety.
| Number | Date | Country | |
|---|---|---|---|
| 63176787 | Apr 2021 | US |
