Image compression/decompression in a computer vision system

Abstract

A computer vision system is provided that includes a camera capture component configured to capture an image from a camera, a memory, and an image compression decompression engine (ICDE) coupled to the memory and configured to receive each line of the image, and compress each line to generate a compressed bit stream. To compress a line, the ICDE is configured to divide the line into compression units, and compress each compression unit, wherein to compress a compression unit, the ICDE is configured to perform delta prediction on the compression unit to generate a delta predicted compression unit, compress the delta predicted compression unit using exponential Golomb coding to generate a compressed delta predicted compression unit, and add the compressed delta predicted compression unit to the compressed bit stream.

Description

BACKGROUND OF THE DISCLOSURE

Field of the Disclosure

Embodiments of the present disclosure generally relate to computer vision systems, and more specifically relate to image compression/decompression in such systems.

Description of the Related Art

Advanced driver assistance systems (ADAS) are designed to increase the situational awareness of a driver and road safety by providing essential information, warnings, and/or automatic intervention to reduce the possibility/severity of an accident. Depending on the safety aspects supported, such systems can be based upon one or more in-vehicle sensor systems such as radar, camera, light detection and range (LIDAR), and ultrasound based systems. Additionally, such systems can integrate and use external information sources such as global positioning systems, car data networks, and vehicle-to-vehicle or vehicle-to-infrastructure communication systems to efficiently and accurately achieve desired goals. Camera based ADAS is emerging as a key differentiator by original equipment manufacturers (OEM). Camera based ADAS uses various computer vision (CV) and machine learning (ML) technologies to perform real-time driving situation analysis and ultimately may use this information to, for example, warn the driver of possible danger or to autonomously intervene to mitigate danger.

The CV and ML techniques along with supporting systems such as image signal processing systems have a very high data bandwidth requirement in which multiple data transfers are needed in real time to meet hard/soft deadlines for availability of the data. For example, tens of Gigabytes of image and signal data may need to be stored and transferred every second between compute units on an embedded system-on-a-chip (SOC) and between the SOC and external memory. Solutions to reduce the memory bandwidth requirement are desirable.

SUMMARY

Embodiments of the present disclosure relate to methods and apparatus for image compression and decompression in a computer vision system. In one aspect, a computer vision system is provided that includes a camera capture component configured to capture an image from a camera, a memory, and an image compression decompression engine (ICDE) coupled to the memory and configured to receive each line of the image, and compress each line to generate a compressed bit stream. To compress a line, the ICDE is configured to divide the line into compression units, and compress each compression unit, wherein to compress a compression unit, the ICDE is configured to perform delta prediction on the compression unit to generate a delta predicted compression unit, compress the delta predicted compression unit using exponential Golomb coding to generate a compressed delta predicted compression unit, and add the compressed delta predicted compression unit to the compressed bit stream.

In one aspect, a method of image compression in a computer vision system is provided that includes receiving a line of an image, and compressing the line to generate a compressed bit stream by dividing the line into compression units, and compressing each compression unit by performing delta prediction on the compression unit to generate a delta predicted compression unit, compressing the delta predicted compression unit using exponential Golomb coding to generate a compressed delta predicted compression unit, and adding the compressed delta predicted compression unit to the compressed bit stream.

BRIEF DESCRIPTION OF THE DRAWINGS

Particular embodiments will now be described, by way of example only, and with reference to the accompanying drawings:

FIG. 1 is a high level block diagram of an example multiprocessor system-on-a-chip (SOC);

FIG. 2 and FIG. 3 are flow diagrams of methods for compressing a line of image data; and

FIG. 4, FIG. 5, FIG. 6, and FIG. 7 are example formats for storing compressed image data.

DETAILED DESCRIPTION OF EMBODIMENTS OF THE DISCLOSURE

Specific embodiments of the disclosure will now be described in detail with reference to the accompanying figures. Like elements in the various figures are denoted by like reference numerals for consistency.

Embodiments of the disclosure provide for image compression and decompression in an embedded computer vision system that reduces the memory bandwidth needed to transfer image data between a system-on-a-chip (SOC) and external memory. In general, the image compression exploits the spatial redundancy of image pixel data in the horizontal direction to generate a compressed representation of the pixel data that reduces the amount of data transferred between compute units on the SOC and the external memory. In various embodiments, the pixel data compression may be performed using a combination of delta prediction, linear quantization, exponential Golomb coding, rate distortion optimization, and/or rate control techniques.

More specifically, each line of an image is divided into compression units, i.e., segments of consecutive pixels, that are compressed and stored in the external memory. In some embodiments, the compression units are losslessly compressed using a combination of delta prediction and exponential Golomb coding. In some embodiments, the compression units are compressed using a hybrid compression technique combining both lossy and lossless compression. The lossy compression uses a combination of quantization, delta prediction, exponential Golomb coding, and rate distortion optimization. In some such embodiments, rate control techniques are also applied. Further, in some embodiments, both lossless and hybrid compression are provided. In addition, in some embodiments, the compressed data is stored in external memory in a format that permits random access retrieval at the line level. In some such embodiments, the format also permits random access retrieval at the compression unit level within a line.

Embodiments are described herein with a focus on the compression protocol. One of ordinary skill in the art, having benefit of this description, will understand how to reverse the compression to perform the decompression.

FIG. 1 is a high level block diagram of an example multiprocessor system-on-a-chip (SOC) 100 configured to support computer vision processing in a camera-based ADAS. The SOC 100 includes dual general purpose processors (GPP) 102, dual digital signal processors (DSP) 104, a vision processor 106, and various hardware accelerators (HWAs) 112 coupled via a high speed interconnect 122. The SOC 100 further includes a direct memory access (DMA) component 108, a camera capture component 110 that may be coupled to an external camera 124, a display management component 114, on-chip random access memory (RAM) 116, e.g., a computer readable medium, and an external memory interface (EMIF) 120 coupled to an external RAM 126, all of which are coupled to the processors 102, 104, the vision processor 106, and the HWAs 112 via the interconnect 122. The camera 124 may be coupled to the SOC 100 via an analog connection, an Ethernet cable, a low-voltage differential signaling (LVDS) cable, or other suitable wired or wireless connection. In addition, the SOC 100 may include a safety component (not shown) that includes safety related functionality to enable compliance with automotive safety requirements. Such functionality may include support for CRC (cyclic redundancy check) of data, clock comparator for drift detection, error signaling, windowed watch-dog timer, and self testing of the SOC for damage and failures.

The GPP 102 hosts the operating system and provides overall control of the operation of the SOC 100 including scheduling of the processing performed by the HWAs 112. The DSP 104 provides support for computer vision processing such as object detection and classification, which may use the output of one or more of the HWAs 112. The vision processor 106 is a vector processor tuned for computer vision processing such as gradient computation, orientation binning, histogram normalization, etc. The HWAs 112 may include, for example, a optical flow accelerator, a vision preprocessing accelerator, and/or an image scaling accelerator.

The external RAM 126 stores image data that may be accessed by various compute units of the SOC 100, i.e., the processors 102, 104, the vision processor 106, and/or the HWAs 112, via an image compression decompression engine (ICDE) 118. To reduce the bandwidth need to access the image data, the image data is compressed by the ICDE 118 prior to storage in the external RAM 126 and is decompressed by the ICDE 118 when read from the external RAM 126. Techniques that may be implemented by the ICDE 118 to compress the image data are described in reference to FIG. 2 and FIG. 3 herein. Formats for storing the compressed data are described in reference to FIG. 4, FIG. 5, FIG. 6, and FIG. 7 herein.

FIG. 2 and FIG. 3 are flow diagrams for methods for compressing a line of image data, one or more or which may be implemented, for example, by the ICDE 118 of FIG. 1. A method may be repeated for each line of an image to compress the entire image. In these methods, a line of image data is divided into multiple compression units, where a compression unit is a segment of consecutive pixels, and each compression unit is compressed separately. The size of a compression unit is implementation dependent and may be any suitable size. For example, the compression unit size may be chosen to minimize read and write overhead to memory, e.g., the external RAM 126.

FIG. 2 is a flow diagram of a method for lossless compression of a line of image data, i.e., the quality of the image data is not affected by the compression. Initially, the line of image data is received 200 and an uncompressed bit stream of the line is generated 202. The line is then divided into compression units and each compression unit is compressed and added to a compressed line level bit stream 204-212. To compress a compression unit, delta prediction is preformed 204 on the pixel data in the compression unit to generate a delta predicted compression unit. That is, in the delta predicted compression unit, the first value is the same as the first value in the original compression unit. Each subsequent value in the delta predicted compression unit is the difference or delta between the corresponding value in the original compression unit and the previous value in the original compression unit. For example, if the compression unit includes the values 100, 107, 106, 101, 105, then the delta predicted compression unit would include the values 100, 7, −1, −5, 4.

The sign bits are extracted 206 from the values in the delta predicted compression unit and the residue is compressed 208 using exponential Golomb (Exp-Golomb) coding. Any suitable order k Exp-Golomb code may be used. In some embodiments, several values of k are evaluated to find the best value for k for the compression unit, i.e., the value that yields the smallest number of bits. The sign bits and the Exp-Golomb coded compression unit are then appended 210 to the compressed line level bit stream. In some embodiments, the sign bits are stored as uncompressed data in the bit stream immediately preceding the Exp-Golomb coded compression unit. The compressed line level bit stream may be formatted to allow for random access at the compression unit level. Example formats for the compressed line level bit stream are described herein in reference to FIG. 4 and FIG. 6. After all of the compression units of the image line are compressed 212, the smallest of the uncompressed bit stream and the compressed bit stream is stored 214 in external memory. The bit stream is stored at a predetermined offset from the previously stored bit stream to permit random access at the line level.

FIG. 3 is a flow diagram of a method for hybrid compression of a line of image data. In this method, a target bit rate is used, i.e., each compression unit is compressed such that the size of the compressed compression unit is less than or equal to a specified number of bits. Any suitable target bit rate may be used and the target bit rate may be user specified. In this method, a compression unit may be losslessly compressed or may be lossy compressed, i.e., the quality of the image data is reduced by the compression.

Initially, the line of image data is received 300 and a quantization parameter is initialized 302 to one, i.e., no quantization. The line is then divided into compression units and each compression unit is compressed and added to a compressed line level bit stream 304-314. To compress a compression unit, the compression unit is quantized 304 according to the quantization parameter and delta prediction is performed 204 on the pixel data in the compression unit to generate a delta predicted compression unit. Delta prediction is previously described herein.

The sign bits are extracted 308 from the values in the delta predicted compression unit and the residue is compressed 310 using exponential Golomb (Exp-Golomb) coding. Any suitable order k Exp-Golomb code may be used. In some embodiments, several values of k are evaluated to find the best value for k for the compression unit, i.e., the value that yields the smallest number of bits. The size of the combination of the sign bits and the Exp-Golomb coded compression unit is then compared 312 to the target bit rate. If the size is less than or equal to the target bit rate, the sign bits and the coded compression unit are appended 314 to the compressed line level bit stream. In some embodiments, the sign bits are stored as uncompressed data in the bit stream immediately preceding the Exp-Golomb coded compression unit. The compressed line level bit stream may be formatted to allow for random access at the line level and/or to allow for random access at the compression unit level. Example formats for the compressed line level bit stream are described herein in reference to FIG. 5 and FIG. 7. Note that if the quantization parameter is one at this point, the coding of the compression unit is lossless.

If the size is not less than or equal to the target bit rate, then the quantization parameter is increased, e.g., incremented by one, and the compression process 304-312 is repeated. After all of the compression units of the image line are compressed 316, the compressed bit stream is stored 316 in external memory. The bit stream is stored at a predetermined offset from the previously stored bit stream to permit random access at the line level.

In some embodiments, the method of FIG. 3 is modified such that rather than requiring a fixed target bit rate for each compression unit in a line, a rate control scheme is used in which if fewer bits than the target bit rate are needed to compress a compression unit, the extra bits are added to the target bit rate for the next compression unit such that, if needed, the number of bits used to code the next compression unit may use the target bit rate plus the extra bits.

FIG. 4 illustrates an example format for storing a lossless compressed line of image data such that each line of the compressed image can be randomly accessed. Line 400 illustrates the generic format, line 402 illustrates the format when the line is compressed, and line 404 illustrates the format when the line is not compressed. Referring first to line 400, an initial bit d in the line indicates whether or not the line is compressed. In this example, if the value of d is zero, then the line is not compressed; if the value is one, then the line is compressed. One of ordinary skill in the art will understand embodiments in which a value of zero indicates compression and a value of one indicates no compression. In the compressed format 402, this bit is followed by the number of bytes LL in the line. Including the number of bytes LL in the compressed bit stream informs the decoder of the size of the data to be read from memory. Further, if the compression units CU-N are compressed as shown in line 402, the bits of each compression unit are preceded by the order k used for the Exp-Golomb coding of the compression unit. In some embodiments, in the compressed format 402, the sign bit is included in each compressed compression unit. In some embodiments, the sign bits of the compressed compression units are stored separately from the compressed compression units and the deltas for each compressed compression unit are the absolute values of the deltas.

FIG. 5 illustrates an example format for storing a hybrid compressed line of image data such that each line of the compressed image can be randomly accessed. Line 500 illustrates the generic format, line 502 illustrates the format with an example combination of lossless and lossy compressed compression units, line 504 illustrates the format of a lossy compressed compression unit, and line 506 illustrates the format of a lossless compressed compression unit. As shown in line 500 and line 502, the initial field in the line contains the number of bytes LL in the line. Including the number of bytes LL in the bit stream informs the decoder of the size of the data to be read from memory. As shown in line 504 and 506, each compressed compression unit begins with a bit indicating whether the compression unit is lossless encoded or lossy encoded.

If the compression unit is lossless encoded, then, as illustrated in line 506, the next field in the compressed compression unit is the order k used for the Exp-Golomb coding of the compression unit and the remainder of the compressed compression unit is the Exp-Golomb coded deltas of the compression unit. If the compression unit is lossy encoded, then, as illustrated in line 504, the next field in the compressed compression unit is the quantization factor used to quantize the pixel values in the compression unit. The quantization factor is followed by the order k used for the Exp-Golomb coding of the compression unit and the remainder of the compressed compression unit is the Exp-Golomb coded deltas of the quantized compression unit. In some embodiments, the sign bit is included in each compressed compression unit. In some embodiments, the sign bits of the compressed compression units are stored separately from the compressed compression units and the deltas for each compressed compression unit are the absolute values of the deltas.

FIG. 6 illustrates an example format for storing a lossless compressed line of image data such that each compression unit of the compressed image can be randomly accessed. In this example, the offset 0(N) of each compressed compression unit in the line relative to the previous coding unit is stored prior to the compressed compression units. In this example, the offset for the first compressed compression unit is the number of bits or bytes needed to represent the value of d and all of the offsets. The other offsets are offsets relative to the previous coding unit, i.e., are the number of bits in the previous compressed compression unit, referred to as CUL (compression unit length) in the figure. In some embodiments, the sign bit is included in each compressed compression unit. In some embodiments, the sign bits of the compressed compression units are stored separately from the compressed compression units and the deltas for each compressed compression unit are the absolute values of the deltas.

FIG. 7 illustrates an example format for storing a hybrid compressed line of image data such that each compression unit of the compressed image can be randomly accessed. In this example, the offset 0(N) of each compressed compression unit in the line relative to the previous compressed compression unit is stored prior to the compressed compression units. In this example, the offset for the first compressed compression unit is the number of bits or bytes needed to represent the value of d and all of the offsets. The other offsets are offsets from the previous compressed compression unit, i.e., are the number of bits in the previous compressed compression unit, referred to as CUL (compression unit length) in the figure. In some embodiments, the sign bit is included in each compressed compression unit. In some embodiments, the sign bits of the compressed compression units are stored separately from the compressed compression units and the deltas for each compressed compression unit are the absolute values of the deltas.

Other Embodiments

While the disclosure has been described with respect to a limited number of embodiments, those skilled in the art, having benefit of this disclosure, will appreciate that other embodiments can be devised which do not depart from the scope of the disclosure as disclosed herein.

For example, embodiments have been described herein in reference to ADAS. One of ordinary skill in the art will understand embodiments for computer vision applications other than ADAS, such as, for example, industrial applications, robotics, and drones.

In another example, embodiments have been described herein in which an SOC includes a single ICDE. One of ordinary skill in the art will understand embodiments including more than one ICDE.

Although method steps may be presented and described herein in a sequential fashion, one or more of the steps shown in the figures and described herein may be performed concurrently, may be combined, and/or may be performed in a different order than the order shown in the figures and/or described herein. Accordingly, embodiments should not be considered limited to the specific ordering of steps shown in the figures and/or described herein.

Certain terms are used throughout the description and the claims to refer to particular system components. As one skilled in the art will appreciate, components in systems may be referred to by different names and/or may be combined in ways not shown herein without departing from the described functionality. This document does not intend to distinguish between components that differ in name but not function. In the description and in the claims, the terms “including” and “comprising” are used in an open-ended fashion, and thus should be interpreted to mean “including, but not limited to . . . .” Also, the term “couple” and derivatives thereof are intended to mean an indirect, direct, optical, and/or wireless electrical connection. Thus, if a first device couples to a second device, that connection may be through a direct electrical connection, through an indirect electrical connection via other devices and connections, through an optical electrical connection, and/or through a wireless electrical connection, for example.

It is therefore contemplated that the appended claims will cover any such modifications of the embodiments as fall within the true scope of the disclosure.

Claims

1. A system comprising: a memory; andan image compression decompression engine (ICDE) coupled to the memory and configured to: receive a line of pixels of an image; divide the line into a compression units, wherein a compression unit is a segment of consecutive pixels; perform delta prediction on each compression unit to generate respective delta predicted compression units;compress each delta predicted compression unit using exponential Golomb coding to generate respective compressed delta predicted compression units; andadd the compressed delta predicted compression units to a compressed bit stream.

2. The system of claim 1, wherein to compress each delta predicted compression unit, the ICDE is further configured to determine a best order k for the exponential Golomb coding of the delta predicted compression unit, wherein a plurality of values of k are evaluated to find a value that yields a smallest number of bits in the respective compressed delta predicted compression unit.

3. The system of claim 1, wherein the ICDE is further configured to: generate an uncompressed bit stream for the line; andstore a smaller of the uncompressed bit stream for the line and the compressed bit stream for the line in the memory.

4. The system of claim 3, wherein the ICDE is further configured to store a bit stream selected from an uncompressed bit stream and a compressed bit stream in the memory for the line of the image, wherein the bit stream is stored at a predetermined offset from a previously stored bit stream to permit random access to each bit stream.

5. The system of claim 1, wherein the ICDE is further configured to store the compressed bit stream in the memory, wherein each compressed delta predicted compression unit in the compressed bit stream can be randomly accessed.

6. The system of claim 5, wherein the compressed bit stream comprises a plurality of offsets each corresponding to one of a plurality of compressed delta predicted compression units, wherein each offset indicates a location of the corresponding compressed delta predicted compression unit in the compressed bit stream relative to a previous compressed delta predicted compression unit.

7. The system of claim 1, wherein the ICDE is further configured to quantize each compression unit before performance of delta prediction.

8. The system of claim 7, wherein the ICDE is further configured to: compare a size of each compressed delta predicted compression unit to a target bit rate; andrepeat the quantization, performance of delta prediction and compression of the delta predicted compression unit when the size is greater than the target bit rate, wherein a quantization parameter for the quantization is increased.

9. The system of claim 8, wherein the ICDE is further configured to increase the target bit rate for a next compression unit when the size is less than the target bit rate, wherein the target bit rate is increased by a number of bits given by a difference between the target bit rate and the size.

10. The system of claim 8, wherein when the quantization parameter used to quantize a compression unit has a value of one and the size of the respective compressed delta predicted compression unit is not greater than the target bit rate, compression of the respective compressed delta predicted compression unit is lossless and otherwise the compression is lossy.

11. The system of claim 1, wherein the ICDE is comprised in a system-on-a-chip (SOC) and the memory is external to the SOC.

12. A method comprising: receiving a line of pixels of an image;dividing the line into a compression units, wherein a compression unit is a segment of consecutive pixels; performing delta prediction on each compression unit to generate respective delta predicted compression units;compressing each delta predicted compression unit using exponential Golomb coding to generate respective compressed delta predicted compression units; andadding the compressed delta predicted compression units to the compressed bit stream.

13. The method of claim 12, wherein compressing each delta predicted compression unit further comprises determining a best order k for the exponential Golomb coding of the delta predicted compression unit by evaluating a plurality of values of k to find a value that yields a smallest number of bits in the respecive compressed delta predicted compression unit.

14. The method of claim 12, further comprising: generating an uncompressed bit stream of the line; andstoring a smaller of the uncompressed bit stream and the compressed bit stream in memory.

15. The method of claim 14, wherein a bit stream selected from an uncompressed bit stream and a compressed bit stream is stored in memory for the line of the image, wherein the bit stream is stored at a predetermined offset from a previously stored bit stream to permit random access to the bit stream.

16. The method of claim 12, further comprising storing the compressed bit stream in memory, wherein each compressed delta predicted compression unit in the compressed bit stream can be randomly accessed.

17. The method of claim 16, wherein the compressed bit stream comprises a plurality of offsets each corresponding to one of a plurality of compressed delta predicted compression units, wherein each offset indicates a location of the corresponding compressed delta predicted compression unit in the compressed bit stream relative to a previous compressed delta predicted compression unit.

18. The method of claim 12, further comprising quantizing each compression unit before performing delta prediction.

19. The method of claim 18, further comprising: comparing a size of each compressed delta predicted compression unit to a target bit rate; andrepeating the quantizing, performing delta prediction and compressing the delta predicted compression unit when the size is greater than the target bit rate, wherein a quantization parameter for the quantizing is increased.

20. The method of claim 19, wherein when the size is less than the target bit rate, the target bit rate for a next compression unit is increased by a number of bits given by a difference between the target bit rate and the size.

21. The method of claim 19, wherein when the quantization parameter used to quantize a compression unit has a value of one and the size of the respective compressed delta predicted compression unit is not greater than the target bit rate, compression of the respective compressed delta predicted compression unit is lossless and otherwise the compression is lossy.