Efficient storage of fingerprints

Abstract

A robust fingerprinting system is disclosed. Such a system can recognize unknown multimedia content (U(t)) by extracting a fingerprint (a series of hash words) from said content, and searching a resembling fingerprint in a database in which fingerprints of a plurality of known contents (K(t)) are stored. In order to more efficiently store the fingerprints in the database and to speed up the search, the hash words (H(n)) of known signals (K(t)) are sub-sampled (13) by a factor M prior to storage in the database (14). The hash words (H(n)) of unknown signals (U(t)) are divided (16) into M interleaved sub-series (H0(n) . . . HM−1(n)). The interleaved sub-series are selectively (17) applied to the database (14) under the control of a computer (15). If only one of the sub-series sufficiently matches a stored fingerprint, the signal is identified.

Description

FIELD OF THE INVENTION

The invention relates to a method and arrangement for storing fingerprints identifying audio-visual media signals in a database. The invention also relates to a method and arrangement for identifying an unknown audio-visual media signal.

BACKGROUND OF THE INVENTION

A fingerprint (in literature also referred to as signature or hash) is a digital summary of an information signal. In cryptography, hashes have been used for a long time to verify correct reception of large files. Recently, the concept of hashing has been introduced to identify multi-media content. Unknown content such as an audio or video clip is recognized by comparing a fingerprint extracted from said clip with a collection of fingerprints stored in a database. In contrast with a cryptographic hash, which is extremely fragile (flipping a single bit in the large file will result in a completely different hash), a fingerprint extracted from audio-visual content must be robust. To a large extent, it must be invariant to processing such as compression or decompression, A/D or D/A conversion.

A prior-art fingerprinting system is disclosed in Haitsma et al.: Robust Hashing for Content Identification, published at the Content-Based Multimedia Indexing (CBMI) conference in Brescia (Italy), 2001. As described in this article, the fingerprint is derived from a perceptually essential property of the content, viz. the distribution of energy in bands of the audio frequency spectrum. For video signals, the distribution of luminance levels in video images has been proposed to constitute the basis for a robust fingerprint.

A fingerprint is created by dividing the signal into a series of (possibly overlapping) frames, and extracting a hash word representing the perceptually essential property of the signal within each frame to obtain a respective series of hash words. In order to identify an unknown clip, the database receives the series of hash words concerned, and searches the most similar stored series of hash words. Similarity is measured by determining how many bits of the series match a series of hash words in the database. If the BER (Bit Error Rate, the percentage of the non-matching bits) is below a certain threshold, the clip is identified as the song or movie from which the most similar series of hash words in the database originates.

A problem of the prior-art fingerprinting method is the size of the database. In the Haitsma et al. article, the audio signal is divided into frames of 0.4 seconds with an overlap of 31/32. This results in a new frame every 11.6 ms (=0.4/32). For every frame, a 32-bit hash word is extracted. Accordingly, a 5-minute song needs approximately 100 kbytes, viz. 5 (minutes)×60 (seconds)×4 (bytes per hash word)/0.0116 (seconds per hash word). Needless to say that the database must have a huge capacity to allow recognition of a large repertoire of songs. Similar considerations apply to video fingerprinting systems.

OBJECT AND SUMMARY OF THE INVENTION

It is an object of the invention to provide a method and system for storing fingerprints in a database, which alleviates the above-mentioned problem. It is also an object of the invention to provide a method and system for identifying an unknown audio-visual signal in such a database.

To this end, the invention provides a method for storing fingerprints in a database as defined in independent claim 1. The method differs from the prior art in that only a sub-sampled sequence of hash words (i.e. one out of every M hash word) is stored in the database. The word “sequence” is used in this claim to refer to a full-length signal (song or movie). A storage reduction by a factor M is achieved.

A method of identifying an unknown audio-visual signal in such a database is defined in independent claim 4. As there is uncertainty as to which of M possible sub-sampled sequences of hash words is stored in the database, a full (i.e. not sub-sampled) series of hash words is extracted from the unknown clip in accordance with this method. The word “series” is used here to refer to a possibly short segment or clip of the unknown signal. Interleaved sub-series of hash words are now successively applied to the database for matching with the sub-sampled sequences stored therein. If at least one of the applied sub-series has a BER below a certain threshold, the signal is identified.

It is achieved with the invention that the storage requirements are reduced (by a factor M), while the robustness and the reliability of the prior-art identification method are maintained.

Further advantageous embodiments of the methods are defined in the dependent claims.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows a schematic diagram of an arrangement for storing and identifying fingerprints of audio-visual media signals in a database in accordance with the invention.

FIG. 2 is a diagram to illustrate a first operational mode of the arrangement which is shown in FIG. 1.

FIG. 3 is a diagram to illustrate a second operational mode of the arrangement which is shown in FIG. 1.

FIG. 4 is a flow chart of operational steps performed by a computer which is shown in FIG. 1.

DESCRIPTION OF EMBODIMENTS

The invention will be described for audio signals. FIG. 1 shows a schematic diagram of an arrangement in accordance with the invention. The arrangement is used for storing fingerprints of known audio signals in a database (first operational mode), as well as for identifying an unknown audio signal (second operational mode).

The first operational mode (storage) of the arrangement will be described first. In this mode, the arrangement receives a fall-length music song K(t). The signal is divided, in a framing circuit 11, into time intervals or frames F(n) having a length of approximately 0.4 seconds and weighted by a Hanning window with an overlap of 31/32. The overlap is used to introduce a large correlation between subsequent frames. For audio signals, this is a prerequisite because the framing applied to unknown signals to be recognized may be different.

The framing circuit 11 generates a new frame every 11.6 ms (=0.4/32). A hash extracting circuit 12 generates a 32-bit hash word H(n) for every frame. A practical embodiment of such a hash extracting circuit is described in the Haitsma et al. article referred to in the chapter Background of the Invention. Briefly summarized, the circuit divides the frequency spectrum of each audio signal frame into frequency bands and produces for each band a hash bit indicating whether the energy in said band is above or below a given threshold. FIG. 2 shows a sequence of hash words 21 thus obtained.

In accordance with the invention, the sequence of hash words is sub-sampled by a factor M by a sub-sampler 13, which produces a sub-sequence H′(n). The sub-sequence of hash words, along with identification data such as title of the song, name of the artist, etc., constitutes a fingerprint of the known music song. Such a fingerprint is shown in FIG. 2, where numeral 22 denotes the sub-sequence of hash words, and numeral 23 denotes title, artist, etc., identifying the song. The fingerprint is stored in a database 14 under the control of a computer 15. In this example, where a sub-sampling factor M=4 has been used by way of example, a 5-minute song requires approximately 6,000×32 bits storage capacity. This is a saving of 75% as compared with the prior-art system where sub-sampling is not applied. In practice, the storage operation described above is performed for a huge number of known music songs. It will be appreciated that the order of the operations of hash word extraction (12) and sub-sampling (13) maybe reversed.

The second operational mode (identification) of the arrangement will now be described. In this mode, the arrangement receives a part (say, 3 seconds) of an unknown song. i.e. an audio clip U(t). The clip is processed by a similar (or the same) framing circuit 11 and hash extracting circuit 12 as described above. The hash extraction circuit 13 extracts a full hash block (no sub-sampling) of the clip. For a 3-second clip, this operation yields a series of approximately 256 hash words H(n). Such a series of hash words representing the unknown audio clip is also referred to as hash block. In an alternative embodiment, the hash block has been extracted by a remote station and is merely received by the arrangement.

The hash block is applied to an interleaving circuit 16, which divides it into M interleaved sub-series or sub-blocks H₀(n), H₁(n), . . . H_M−1(n), where M is the same integer as used in the sub-sampler 13 described above. FIG. 3 illustrates the interleaving process for M=4. In this Figure, numeral 31 denotes successive hash words of the hash block, numeral 32 denotes sub-block H₀(n), numeral 33 denotes sub-block H₁(n), and numeral 34 denotes sub-block H_M−1(n).

The sub-blocks are applied to respective inputs of a selection circuit 17. Under the control of the computer 15, the sub-blocks H₀(n), H₁(n), . . . H_M−1(n) are successively applied to the database 14 for identification. If a series of hash words is found in the database, for which the bit error rate BER (i.e. the percentage of non-matching bits between said series and the applied sub-block) is below a certain threshold, the fingerprint comprising said series of hash words identifies the unknown audio clip.

FIG. 4 shows a flow chart of this identification process which is performed by the computer 15. In a step 41, an index m obtains an initial value 0. The index m is applied to the selection circuit 17 so that the first interleaved sub-block H₀(n) of hash words is selected for identification. In a step 42, the selected sub-block H_m(n) is applied to the database. In a step 43, it is checked whether a resembling series of hash words has been found in the database. The word “resembling” is understood to refer to the series of hash words having the lowest BER provided that said BER, is less than a given threshold T. An actual example of a strategy of searching the most resembling series of hash words in the database is disclosed in the Haitsma et al. article mentioned before. Advantageous embodiments of search strategies are also proposed in Applicant's pending unpublished European patent applications 01200505.4 (PHNL010110) and 01202720.7 (PHNL010510).

If the BER is below the threshold, the audio clip has been identified. The title and performer of the song as stored in the database (23 in FIG. 2) are then communicated to the user in a step 44. If that is not the case, the index m is incremented (step 45) so that another one of the interleaved sub-blocks is applied to the database. If all M interleaved sub-blocks have been searched without success (step 46), the audio clip could not be identified. This outcome is communicated to the user in a step 47.

It is achieved with the invention that the database capacity is reduced by a factor M. It should be noted that the same reduction can effectively be achieved by choosing a different frame overlap, viz. ⅞ in the present example. This is true as far as the first operational mode (storage) is concerned. However, if the same overlap of ⅞ without interleaving was chosen in the identification process, the robustness and reliability of the identification would be seriously affected. The invention resides in the concept of interleaving in the second operational mode (identification). It is achieved thereby that at least one of the interleaved sub-blocks is derived from a series of frames that substantially matches (in time) the series of frames from which the stored hash words have been derived. The identification process in accordance with the invention yields substantially the same robustness and reliability as the prior-art (non-interleaving) method with an overlap of 31/32. A mathematical background thereof will now be given.

When a sub-sampling with a factor M is applied and if the bits in a hash block are random i.i.d. (independent and identically distributed), the standard deviation of the BER increases by a factor √M. This implies that either the robustness and/or the reliability is/are affected. If the threshold on the BER is kept the same, then the robustness is unaffected but the reliability decreases. If on the other side the threshold is decreased by an appropriate amount, then the reliability stays the same but the robustness decreases.

However, the bits in a hash block of an audio-visual media signal have a large correlation along the time axis, which is introduced by the large overlap of the framing and inherent correlation in music. Therefore, the standard deviation s does not increase by the factor √M when sub-sampling with the factor M is applied. Experiments have shown that, for small values of M, the standard deviation does not even increase significantly at all. In a practical system without sub-sampling, the threshold on BER is set to 0.35. If sub-sampling by a factor M=4 is applied, then the threshold has only to be lowered to 0.342. Therefore, the decrease of robustness is insignificant, whilst the needed storage in the database has been decreased by a factor of 4. Furthermore, the time needed to search a hash database will decrease simply because there are 4 times fewer hash values in the database.

The search speed can even be further increased by refraining from applying a further sub-block to the database if one of the sub-blocks (generally the first) appears to have a BER which is larger than a further threshold (which is substantially larger than the threshold T). Because of the large correlation between sub-blocks (due to the frame overlap and inherent correlation in music), it is unlikely that another sub-block will have a significantly lower BER.

A robust fingerprinting system is disclosed. Such a system can recognize unknown multimedia content (U(t)) by extracting a fingerprint (a series of hash words) from said content, and searching a resembling-fingerprint in a database in-which fingerprints of a plurality of known contents (K(t)) are stored. In order to more efficiently store the fingerprints in the database and to speed up the search, the hash words (H(n)) of known signals (K(t)) are sub-sampled (13) by a factor M prior to storage in the database (14). The hash words (H(n)) of unknown signals (U(t)) are divided (16) into M interleaved sub-series (H₀(n) . . . H_M−1(n)). The interleaved sub-series are selectively (17) applied to the database (14) under the control of a computer (15). If only one of the sub-series sufficiently matches a stored fingerprint, the signal is identified.

Claims

1. A method of storing fingerprints identifying audio-visual media signals in a database, the method comprising, for each audio-visual signal: dividing said audio-visual media signal into a sequence of frames;sub-sampling said sequence of frames by a factor M to obtain a sub-sampled sequence of frames, each frame of the sub-sampled sequence of frames overlapping in time with an adjacent frame of the sub-sampled sequence of frames, and the factor M being a positive integer;extracting, for each frame of said sub-sampled sequence of frames, a hash word derived from a perceptually essential property of the signal within said frame, to obtain a respective sub-sampled sequence of hash words, each hash word of the sub-sampled sequence of hash words being nonrandomly positioned within the sub-sampled sequence of hash words; andstoring said sub-sampled sequence of hash words as a fingerprint in said database, the fingerprint being a digital summary of the audio visual media signal.

2. An arrangement to store fingerprints identifying audio-visual media signals (K(t)) in a database, the arrangement comprising: framing means for dividing said audio-visual media signals into a sequence of overlapping frames;sub-sampling means for sub-sampling said sequence of frames by a factor M to obtain a sub-sampled sequence of frames, each frame of the sub-sampled sequence of frames overlapping in time with an adjacent frame of the sub-sampled sequence of frames, and, the factor M being a positive integer;means for extracting, for each frame of said sub-sampled sequence of frames, a hash word (H(n)) derived from a perceptually essential property of the signal within said frame, to obtain a respective sub-sampled sequence of hash words, each hash word of the sub-sampled sequence of hash words being nonrandomly positioned within the sub-sampled sequence of hash words; anda database for storing said sub-sampled sequence of hash words as fingerprint in said database, the fingerprint being a digital summary of the audio visual media signal.

3. A method of identifying an unknown audio-visual media signal, the method comprising: dividing at least a part of the unknown audio-visual media signal into a series of frames;extracting, for each frame, a hash word representing a perceptually essential property of the signal within said frame, to obtain a respective series of hash words;dividing said series of hash words into M interleaved sub-series of hash words, each hash word of each of the M interleaved sub-series of hash words being extracted from a different frame of the series of frames, each frame of the series of frames overlapping in time with an adjacent frame of the series of frames, from which an adjacent hash word of the M interleaved sub-series of hash words is extracted, and each hash word of each of the M interleaved sub-series of hash words being nonrandomly positioned within each of the M interleaved sub-series of hash words;successively applying each of said M interleaved sub-series to a database in which, for a plurality of multi-media signals, a sub-sampled sequence of hash words has been stored; andidentifying the unknown signal as the multi-media signal based on whether a difference between at least a part of the stored sub-sampled sequence of hash words and at least one of the M applied interleaved sub-series of hash words is less than a specific threshold value.

4. An arrangement to identify an unknown audio-visual media signal, the arrangement comprising: framing means for dividing at least a part of the unknown audio-visual media signal (U(t)) into a series of frames;means for extracting, for each frame, a hash word derived from a perceptually essential property of the signal within said frame, to obtain a respective series of hash words;interleaving means for dividing said series of hash words into M interleaved sub-series of hash words, each hash word of each of the M interleaved sub-series of hash words being extracted from a different frame of the series of frames, each frame of the series of frames overlapping in time with an adjacent frame of the series of frames, from which an adjacent hash word of the M interleaved sub-series of hash words is extracted, and each hash word of each of the M interleaved sub-series of hash words being nonrandomly positioned within each of the M interleaved sub-series of hash words;selection means for successively applying each of said M interleaved sub-series to a database in which for a plurality of multi-media signals, a sub-sampled sequence of hash words has been stored; andcomputer means for identifying the unknown signal as the multi-media signal based on whether a difference between at least a part of the stored sub-sampled sequence of hash words and at least one of the M applied interleaved sub-series of hash words is less than a specific threshold value.

5. A method of identifying an unknown audio-visual media signal, the method comprising: receiving, from a remote station, a series of hash words generated by dividing at least a part of the unknown audio-visual media signal into a series of frames, and extracting, for each frame, a hash word based on a perceptually essential property of the signal within said frame;dividing said series of hash words into M interleaved sub-series of hash words, each hash word of each of the M interleaved sub-series of hash words being extracted from a different frame of the series of frames, each frame of the series of frames overlapping in time with an adjacent frame of the series of frames, from which an adjacent hash word of the M interleaved sub-series of hash words is extracted, and each hash word of each of the M interleaved sub-series of hash words being nonrandomly positioned within each of the M interleaved sub-series of hash words;successively applying each of said M interleaved sub-series to a database in which, for a plurality of multi-media signals, a sub-sampled sequence of hash words has been stored; andidentifying the unknown signal as the multi-media signal based on whether a difference between at least a part of the stored sub-sampled sequence of hash words and at least one of the M applied interleaved sub-series of hash words is less than a specific threshold value.

6. A method as claimed in claim 5, wherein each frame of said series of frames are overlapping in time with an adjacent frame.

7. A system to store fingerprints identifying audio-visual media signals (K(t)) in a database, the arrangement comprising: a framing circuit to divide said audio-visual media signals into a sequence of overlapping frames;sub-sampler to sub-sample said sequence of frames by a factor M to obtain a sub-sampled sequence of frames, each frame overlapping in time with an adjacent frame of the sub-sampled sequence of frames, and the factor M being a positive integer;a hash extracting circuit to extract for each frame of said sub-sampled sequence of frames, a hash word (H(n)) derived from a perceptually essential property of the signal within said frame, to obtain a respective sub-sampled sequence of hash words, each hash word of the sub-sampled sequence of hash words being nonrandomly positioned within the sub-sampled sequence of hash words; anda database for storing said sub-sampled sequence of hash words as a fingerprint in said database, the fingerprint being a digital summary of the audio visual media signal.

8. A system to identify an unknown audio-visual media signal, the arrangement comprising: a framing circuit to divide at least a part of the unknown audio-visual media signal (U(t)) into a series of frames;a hash extracting circuit, to extract for each frame, a hash word derived from a perceptually essential property of the signal within said frame, to obtain a respective series of hash words;an interleaving circuit to divide said series of hash words into M interleaved sub-series of hash words, each hash word of each of the M interleaved sub-series of hash words being extracted from a different frame of the series of frames, each frame of the series of frames overlapping in time with an adjacent frame of the series of frames, from which an adjacent hash word of the M interleaved sub-series of hash words is extracted, and each hash word of each of the M interleaved sub-series of hash words being nonrandomly positioned within each of the M interleaved sub-series of hash words;a selection circuit to successively apply each of said M interleaved sub-series to a database in which for a plurality of multi-media signals, a sub-sampled sequence of hash words has been stored; anda computer to identify the unknown signal as the multi-media signal based on whether a difference between at least a part of the stored sub-sampled sequence of hash words and at least one of the M applied interleaved sub-series of hash words is less than a specific threshold value.