482results about How to "Improve decoding speed" patented technology

A method and device for deblocking filtering of a scalable bitstream during decoding is disclosed. According to the invention, the decoding of at least one part of a picture encoded in a base layer and at least one enhancement layer in a scalable bitstream, the bitstream comprising at least one access unit representing the picture, each access unit comprising a plurality of data elements, each data element belonging to a said layer, comprises the reception (600) of at least one data element of said at least one access unit and, if said at least one data element belongs to an access unit of a predetermined type, the decoding (715, 820) of the data of said data element, and the application (720, 825) of a full deblocking filter to at least one part of the decoded data, the full deblocking filter being applied to all of the.

Systems, devices and methods are provided to compress data, and in particular to code and decode data. One aspect of the present subject matter is a data structure. The data structure includes a field representing a decoding structure to decode canonical Huffman encoded data, and a field representing a symbol table. The decoding structure includes a field representing an accelerator table to provide a 2^N-deep direct-index lookup to provide high-frequency symbols for high-frequency data and to provide bracketing indices for low-frequency data. The decoding structure also includes a field for a binary search table to provide a low-frequency symbol index using a binary search bounded by the bracketing indices provided by the accelerator table. The symbol table is adapted to provide a symbol associated with the low-frequency index.

A method is provided for reducing search complexity in a speech recognition system having a fast match, a detailed match, and a language model. Based on at least one predetermined variable, the fast match is optionally employed to generate candidate words and acoustic scores corresponding to the candidate words. The language model is employed to generate language model scores. The acoustic scores are combined with the language model scores and the combined scores are ranked to determine top ranking candidate words to be later processed by the detailed match, when the fast match is employed. The detailed match is employed to generate detailed match scores for the top ranking candidate words.

The invention relates to the field of the language processing, and provides a decoding method based on a deep neural network translation model. The problems of high model training complexity, large training difficulty and slow decoding speed in a machine translation model are solved. The specific implementation way of the method comprises the following steps: performing word segmentation processing on the to-be-translated sentence to obtain source language vocabulary; step two, performing word alignment on the linguistic data in a preset translation model glossary by using an automatic alignment tool so as to obtain a target language word aligned to the source language vocabulary; and step three, determining a target-side dynamic glossary of the to-be-translated sentence based on the target language word obtained in the step two, and using the sentence decoded by using a column searching method as the output of the translation model according to the pre-constructed translation model, wherein the translation model is the deep neural network based on the threshold residual mechanism and the parallel attention mechanism. Through the decoding method disclosed by the invention, the model translation quality is improved, and the mode coding speed is improved.

The invention discloses a two-dimensional recursive network-based recognition method of Chinese text in natural scene images. Firstly, a training sample set is acquired, and a neural network formed bysequentially connecting a deep convolutional network, a two-dimensional recursive network used for encoding, a two-dimensional recursive network used for decoding and a CTC model is trained; test samples are input into the trained deep convolutional network, and feature maps of the test samples are acquired; the feature maps of the test samples are input into the trained two-dimensional recursivenetwork, which is used for encoding, to obtain encoding feature maps of the test samples; the encoding feature maps of the test samples are input into the trained two-dimensional recursive network, which is used for decoding, to obtain a probability result of each commonly used Chinese character in each image of the test samples; and clustering searching processing is carried out, and finally, the overall Chinese text in the test samples is recognized. According to the method of the invention, space/time information and context information of the text images are fully utilized, the text imagepre-segmentation problem can be avoided, and recognition accuracy is improved.

An image is divided into subbands by wavelet transform using the Haar function as the base, and the lowest-frequency LL subband is entirely encoded. LH, HL, and HH subband coefficients which belong to the wavelet decomposition level of each hierarchy are then encoded such that coefficients at the same spatial position are encoded at once. The decoding side first decompresses the lowest-frequency LL subband, and then decodes sets of the LH, HL, and HH coefficients at the same spatial position in the subband of each wavelet decomposition level one by one. The decoding side immediately performs inverse wavelet transform by using the coefficient values, thereby obtaining the LL coefficient value of the next wavelet decomposition level. This makes it possible to sufficiently increase the processing speed even when the wavelet encoding/decoding is performed using a sequential CPU.

The invention discloses software and hardware-decoding general multi-media playing equipment, which comprises a multi-media information acquisition device, a multi-media processing device and a multi-media display device. The general multi-media playing equipment is characterized in that: the multi-media processing device comprises a file analyzing module for analyzing multi-media information according to the types of multi-media files, a decoding module for decoding analyzed multi-media audio/video steaming information and an output module for receiving the multi-media information decoded by the decoding module and outputting the multi-media information to the multi-media display device for displaying. The general multi-media playing equipment can utilize processing capacity of a hardware platform to a maximum degree and improve the processing capacity of multi-media.

The invention provides an n-gram grammar model constructing method for voice identification and a voice identification system. The method comprises: step (101), training is carried out by using a neural network language model to obtain word vectors, and classification and multi-layer screening is carried out on word vectors to obtain parts of speech; step (102), manual marking is expanded by using a direct word frequency statistic method; and when same-kind-word substitution is carried out, direct statistics of 1-to-n-gram grammar combination units changing relative to an original sentence is carried out, thereby obtaining an n-gram grammar model of the expanding part; step (103), manual marking is carried out to generate a preliminary n-gram grammar model, model interpolation is carried out on the preliminary n-gram grammar model and the n-gram grammar model of the expanding part, thereby obtaining a final n-gram grammar model. In addition, the step (101) includes: step (101-1), inputting a mark and a training text; step (101-2), carrying out training by using a neural network language model to obtain corresponding work vectors of words in a dictionary; step (101-3), carrying out word vector classification by using a k mean value method; and step (101-4), carrying out multi-layer screening on the classification result to obtain parts of speech finally.

The invention requests to protect a Chinese speech recognition method combining Transformer and CNN-DFSMN-CTC. The method comprises the steps of: S1, preprocessing a speech signal, and extracting 80-dimensional log mel Fbank features; S2, carrying out convolution on the extracted 80-dimensional Fbank features by using a CNN convolution network; S3, inputting the features into a DFSMN network structure; S4, taking CTC loss as a loss function of an acoustic model, predicting by adopting a Beam search algorithm, and optimizing by using an Adam optimizer; S5, introducing a strong language model Transformer for iterative training until an optimal model structure is achieved; and S6, combining the Transformer with the acoustic model CNN-DFSMN-CTC to carry out adaptation, and carrying out verification on multiple data sets to finally obtain an optimal identification result. According to the method, the recognition accuracy is higher, the decoding speed is higher, the character error rate reaches 11.8% after verification on a plurality of data sets, and the best character error rate reaches 7.8% on an Aidatang data set.

The embodiment of the invention discloses a construction method of check codes with low density, an encoder and a decoder. A message part and a redundant part of the LDPC codes to be constructed are respectively divided into M message subsets and K check subsets and are respectively stored in the different lines of an encoding matrix; a circulating shift matrix is generated according to code length and code rate of the LDPC codes to be constructed, elements in the encoding matrix are carried out the circulating shift according to the circulating shift matrix before carrying out exclusive or plus of the elements in the encoding matrix by lines each time respectively, thereby leading information bits and check bits in the same line in the encoding matrix which participate in the exclusive or plus to be different and further generating the different check bits by the exclusive or plus; therefore, the embodiment of the invention can construct the LDPC codes by the search of the circulating shift matrix, compared with a great number of matrix operations in the prior art, the construction method, the encoder and the decoder of the invention can reduce the complexity for constructing, encoding and decoding the LDPC codes.

A method and apparatus for generating a Joint Photographic Experts Group (JEPG) file achieves fast decoding speed by enabling efficient locating of restart markers at a decoder. The method and apparatus determining a location to insert at least one restart marker by analyzing an image, performing entropy-encoding, by dividing the image into a plurality of segments based on the determined location in order to insert the restart marker, and inserting the restart marker to the determined location in order to insert the reset maker with respect to each of the plurality of entropy-encoded segments. The method performed by the apparatus includes calculating an offset value of the at least one restart marker with respect to each of the plurality of segments, based on a data size of each of the plurality of segments and a data size of the restart marker, calculating an offset value by scan number from a start of image (SOI) marker based on the calculated offset value of the restart marker, in response to the calculated offset value of the restart marker corresponding to a last segment that is entropy-encoded. The method and apparatus further store the calculated offset value by scan number, as additional application data.

The invention belongs to the speech recognition technical filed and relates to a speech recognition method and system. The method includes the following steps that: speech signals are acquired; analog-digital conversion is performed on the speech signals, so that corresponding speech digital signals can be generated; preprocessing is performed on the speech digital signals, and speech feature parameters are extracted according to corresponding preprocessing results, and a time sequence of extracting the speech feature parameters is utilized to construct a corresponding feature sequence; the speech feature parameters are matched with speech models in a template library, and the feature sequence is decoded according to a search algorithm, and therefore, a corresponding recognition result can be generated. According to the speech recognition method and system of the invention, time-domain GFCC (gammatone frequency cepstrum coefficient) features are extracted to replace frequency-domain MFCC (mel frequency cepstrum coefficient) features, and DCT conversion is adopted, and therefore, computation quantity can be reduced, and computation speed and robustness can be improved; and the mechanism of weighted finite state transformation is adopted to construct a decoding model, and smoothing and compression processing of the model is additionally adopted, and therefore, decoding speed can be increased.

The invention discloses a fast multivariate LDPC code decoding method with a low decoding complexity, belonging to the technical field of mobile communication channel coding. According to the multivariate LDPC code decoding method, when storing decoding messages, only the decoding messages corresponding to the location of a non-zero element in a check matrix H are stored, and due to the sparsity of the check matrix H, the decoding method can greatly reduce the storage space occupied by the decoding messages when being implemented by using software or hardware. During decoding, the addressing speed of the decoding messages can be accelerated by using pre-calculated information of the non-zero element in the check matrix H, wherein the information includes the row number, the column number and an element value of the non-zero element; and meanwhile, in a decoding iteration process, fast Fourier transform is introduced to reduce the complexity of a decoding operation when processing a check node message, and thus the decoding speed can be accelerated. The decoding method disclosed by the invention has a relatively low decoding complexity, requires a small storage space, and can also increase the decoding speed to a great extent.

The disclosed segmenting method for Turbo code comprises: deciding whether the transmission block length over the system-set threshold; if equal or less than the threshold, using interlace device length KI for segmentation; or else, using near KI-1 and KI for segmentation. This invention ensures system performance, reduces filling-bit number, and decreases the complexity and time delay.

The invention provides a video decoding data storage method and a calculation method of motion vector data. The video decoding data storage method includes setting a reference frame queue table in a static random access memory, storing basic information of a plurality of reference frames in the reference frame queue table, wherein each unit of the reference frame queue table stores index information corresponding to one reference frame in a decoding buffer area; storing a plurality of groups of frame buffer information corresponding to the index information in the decoding buffer area, wherein each group of frame buffer information contains frame display sequence data and motion vector storage address information; storing the motion vector storage of each reference frame in a dynamic random access memory, wherein the motion vector storage address information is address information of the motion vector data of a coding tree unit of corresponding blocks of the reference frames stored in the dynamic random access memory. The video decoding data storage method and the calculation method of the motion vector data can improve the video decoding efficiency and save hardware expense and bandwidth resources occupied during video decoding.

The invention provides a JPEG image decoding method suitable for a multi-core embedded type platform. The method includes the steps of firstly, segmenting a JPEG image into image blocks in different sizes according to the loading condition of each operation core and distributing the image blocks to the operation cores respectively; then, conducting decoding and zooming on the image blocks distributed to the operation cores in a parallel mode through decoding threads operated in different operation cores and obtaining a decoding result of each image block; finally, conducting splicing on the decoding results of the image blocks to obtain a decoded image. The invention further provides a JPEG image decoder suitable for the multi-core embedded type platform. Through the JPEG image decoding method and decoder suitable for the multi-core embedded type platform, system computing resources can be effectively used, and the image decoding speed can be obviously improved.

A method for realizing Huffman decoding includes receiving data stream being Huffman-coded, intercepting different cipher of data stream and using it as word to be decoded, comparing word to be decoded separately with address indexing table used to indicate code word position in Huffman code table to obtain corresponding Huffman code table address then obtaining data symbol before Huffman table by seeking Huffman table according to Huffman code table address and outputting obtained data symbol out. The device used for realizing said method is also disclosed.

The invention relates to the field of video processing, in particular to video decoding technology, and provides a video processing device and method for decoding the coded data obtained by the block-based predictive coding technology so as to achieve the aim of flexibly setting the type and storage block of a data cache according to the requirement for saving bandwidth of external memory, thereby saving the bandwidth of the external memory and increasing the decoding speed.