WO2022153480A1 - 情報処理装置、情報処理システム、情報処理方法および記録媒体 - Google Patents
情報処理装置、情報処理システム、情報処理方法および記録媒体 Download PDFInfo
- Publication number
- WO2022153480A1 WO2022153480A1 PCT/JP2021/001240 JP2021001240W WO2022153480A1 WO 2022153480 A1 WO2022153480 A1 WO 2022153480A1 JP 2021001240 W JP2021001240 W JP 2021001240W WO 2022153480 A1 WO2022153480 A1 WO 2022153480A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- data
- feature
- unit
- processing
- target
- Prior art date
Links
- 230000010365 information processing Effects 0.000 title claims abstract description 101
- 238000003672 processing method Methods 0.000 title claims description 12
- 238000004891 communication Methods 0.000 claims abstract description 137
- 238000000034 method Methods 0.000 claims abstract description 133
- 230000008569 process Effects 0.000 claims abstract description 101
- 238000012545 processing Methods 0.000 claims description 337
- 238000000605 extraction Methods 0.000 claims description 51
- 230000005540 biological transmission Effects 0.000 claims description 34
- 238000005070 sampling Methods 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 description 56
- 238000010586 diagram Methods 0.000 description 44
- 238000013139 quantization Methods 0.000 description 42
- 238000013500 data storage Methods 0.000 description 34
- 230000008878 coupling Effects 0.000 description 27
- 238000010168 coupling process Methods 0.000 description 27
- 238000005859 coupling reaction Methods 0.000 description 27
- 238000006243 chemical reaction Methods 0.000 description 26
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 description 22
- 238000007781 pre-processing Methods 0.000 description 12
- 238000011144 upstream manufacturing Methods 0.000 description 12
- 238000013527 convolutional neural network Methods 0.000 description 11
- 230000006870 function Effects 0.000 description 11
- 230000006835 compression Effects 0.000 description 9
- 238000007906 compression Methods 0.000 description 9
- 238000012544 monitoring process Methods 0.000 description 7
- 239000000284 extract Substances 0.000 description 6
- 238000013528 artificial neural network Methods 0.000 description 5
- 238000001514 detection method Methods 0.000 description 4
- 238000013136 deep learning model Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000012805 post-processing Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 2
- 230000001537 neural effect Effects 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 230000001629 suppression Effects 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 241000499469 Collimonas pratensis Species 0.000 description 1
- 241000526754 Pannonibacter phragmitetus Species 0.000 description 1
- 230000002411 adverse Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000005549 size reduction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
- G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
- G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
- G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/42—Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/28—Quantising the image, e.g. histogram thresholding for discrimination between background and foreground patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/85—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using pre-processing or post-processing specially adapted for video compression
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
Definitions
- the present invention relates to an information processing device, an information processing system, an information processing method and a recording medium.
- the moving image data when the moving image data is compressed and transmitted, the moving image is restored from the received data, and image recognition is performed on the reproduced image, the moving image data is compressed, the moving image is restored, and the moving image is restored.
- image recognition there may be a delay due to processing time.
- the effect of delay is large.
- QoE Quality Of Experience, experience quality of services, etc.
- An example of an object of the present invention is to provide an information processing device, an information processing system, an information processing method, and a recording medium capable of solving the above-mentioned problems.
- the information processing apparatus has a receiving means for receiving communication data based on the feature data indicating the characteristics of the expression content of the target data, and the feature data based on the received communication data.
- a feature restoration means for restoring the target data a target restoration means for restoring the target data based on the restored feature data, and a recognition means for performing recognition processing for the expression content of the target data based on the restored feature data.
- the information processing system includes a transmitting side device and a receiving side device
- the transmitting side device includes a data acquisition means for acquiring target data and an expression content of the target data.
- the receiving device includes a feature extracting means for calculating feature data indicating features, a communication data generating means for generating communication data based on the feature data, and a transmitting means for transmitting the communication data.
- a recognition means that performs recognition processing for the expression content of the target data based on the feature data, and an output means that outputs information indicating the restored expression content of the target data and the recognition result by the recognition processing.
- the information processing method receives communication data based on the feature data indicating the characteristics of the expression content of the target data, and obtains the feature data based on the received communication data. Restoring, restoring the target data based on the restored feature data, performing recognition processing for the expression content of the target data based on the restored feature data, and restoring the above. It includes outputting information indicating the expression content of the target data and the recognition result by the recognition process.
- the recording medium receives the communication data based on the feature data indicating the features of the expression content of the target data on the computer, and the features based on the received communication data. Restoring the data, restoring the target data based on the restored feature data, performing recognition processing for the expression content of the target data based on the restored feature data, and restoring. It is a recording medium for recording a program for outputting information indicating the expression content of the target data and the recognition result by the recognition process and executing the operation.
- the processing time for restoring the target data and recognizing the expression content of the restored data can be relatively short.
- the information processing system performs transmission / reception of image data and image recognition
- the target of transmission / reception and recognition processing in the following embodiments is not limited to image data, and can be various data that can be compressed and decompressed (restored) hierarchically.
- the information processing system may send and receive voice data and perform voice recognition.
- the information processing system may target the point cloud data output by various measuring devices such as LiDAR (Light Detection And Ranging) for transmission / reception and recognition processing.
- LiDAR Light Detection And Ranging
- FIG. 1 is a schematic block diagram showing a configuration example of an information processing system according to the first embodiment.
- the information processing system 1 includes a transmitting side device 10 and a receiving side device 20.
- the transmission side device 10 includes an image acquisition unit 11, a feature extraction unit 12, a communication data generation unit 13, and a transmission unit 16.
- the communication data generation unit 13 includes a quantization unit 14 and a coding unit 15.
- the receiving device 20 includes a receiving unit 21, a feature restoration unit 22, an acquired image restoration unit 26, a recognition unit 27, and an output unit 28.
- the feature restoration unit 22 includes a decoding unit 23, a dequantization unit 24, and an intermediate feature generation unit 25.
- the information processing system 1 transmits images and recognizes images.
- the transmitting side device 10 acquires an image, converts the acquired image into transmission data such as a bit stream, and transmits the acquired image to the receiving side device 20.
- the receiving side device 20 restores an image from the data received from the transmitting side device 10 and also performs image recognition on the received image.
- the information processing system 1 may be a remote monitoring system such as monitoring of an autonomous driving vehicle.
- the transmitting side device 10 may be installed at a monitoring point, and the receiving side device 20 may be installed at a point away from the transmitting side device 10 such as a data center.
- the receiving device 20 may detect or predict the danger in the autonomous driving vehicle by image recognition and notify the danger.
- the use of the information processing system 1 is not limited to a specific use.
- the receiving side device 20 When transmitting an image from the transmitting side device 10 to the receiving side device 20, feature extraction of the image is performed using the learning model, and feature data indicating the extracted feature is transmitted (data conversion is performed as necessary). It may be. Then, the receiving side device 20 may restore the image based on the received feature data.
- the receiving side device 20 performs image recognition using the intermediate feature data generated in the process of image restoration from the received data. As a result, after restoring the image from the received data, the processing can be performed efficiently in a shorter time than in the case of performing image recognition using the restored image.
- the receiving device 20 corresponds to an example of an information processing device.
- the features of the image may be represented by a vector having a real number as an element. That is, the feature data indicating the features of the image may be shown in the form of a feature vector.
- the feature vector is also called a feature quantity or a feature quantity vector.
- the image acquisition unit 11 acquires an image as image data.
- the image acquisition unit 11 may be provided with an image pickup device such as a still camera or a video camera to capture a moving image or a still image.
- the imaging may be repeated at predetermined time intervals.
- the image pickup device may be configured as a device different from the transmission side device 10, and the image acquisition unit 11 may acquire image data from the image pickup device.
- the image acquisition unit 11 may read the image data from the recording medium on which the image data is recorded. The image acquisition unit 11 outputs the acquired image data to the feature extraction unit 12.
- the data format of the image data acquired by the image acquisition unit 11 is not limited to a specific one.
- the image acquisition unit 11 may acquire image data in the RGB pixel data (RGB Pixel Data) format, but the present invention is not limited to this.
- the RGB pixel data format is an image data format in which values of red, green, and blue are shown for each pixel.
- the image acquired by the image acquisition unit 11 is referred to as an acquired image.
- Image data indicating an acquired image is referred to as acquired image data.
- the acquired image data corresponds to an example of target data.
- the acquired image corresponds to an example of the expression content of the target data.
- the image acquisition unit 11 corresponds to an example of acquisition means.
- the feature extraction unit 12 extracts features of the acquired image and generates feature data.
- the feature data is data representing the visual features of the acquired image.
- "visual” indicates that it is a feature relating to the display content of an image rather than an image format or a file format.
- the feature data may be presented in the form of a real vector.
- the feature extraction unit 12 corresponds to an example of a feature extraction means.
- the feature extraction unit 12 may include a neural network model obtained by using a technique of deep learning.
- the neural network model in that case may be an Invertible Neural Network (INN), which is a neural network capable of mathematically inverse calculation.
- INN Invertible Neural Network
- the configuration of the feature extraction unit 12 is not limited to a specific configuration as long as it can generate feature data that can restore the acquired image.
- Generating feature data is also referred to as extracting features or extracting feature data.
- Generating feature data that indicates the features of an image that represents the content of the image data is also referred to as extracting feature data from the image data.
- a deep learning model using a convolutional neural network capable of inverse calculation is also called an Invertible Deep Convolutional Neural Network Model.
- the inverse operation referred to here is an operation in which input / output is reversed from the original operation. That is, in the inverse operation, when the output value in the original operation becomes the input value in the inverse operation, the same value as the input value in the original operation is output.
- FIG. 2 is a schematic block diagram showing a configuration example of the feature extraction unit 12.
- the feature extraction unit 12 includes a preprocessing unit 111, a processing stage unit 112, and a channel dividing unit 113.
- the feature extraction unit 12 includes three processing stage units 112 and two channel division units 113. These are connected in series in an arrangement in which one channel dividing unit 113 is provided between each of the two processing stage units 112, and further connected in series to the preprocessing unit 111.
- reference numerals 112-1, 112-2, and 112-3 are attached in order from the upstream side to the downstream side of the data flow.
- reference numerals 113-1 and 113-2 are attached in order from the upstream side to the downstream side of the data flow.
- the number of processing stage units 112 included in the feature extraction unit 12 may be one or more.
- the number of channel division units 113 included in the feature extraction unit 12 may be one less than the number of processing stage units 112.
- the pre-processing unit 111 performs pre-processing for feature extraction on the image data output by the image acquisition unit 11.
- the preprocessing unit 111 may process the image so that the image size of the image data output by the image acquisition unit 11 matches the image size received by the neural network constituting the feature extraction unit 12. .. Further, the preprocessing unit 111 may apply the image filter to the image data output by the image acquisition unit 11, such as a noise filter when the image output by the image acquisition unit contains a lot of noise.
- the feature extraction unit 12 may not include the preprocessing unit 111. That is, preprocessing by the preprocessing unit 111 is not essential.
- Each output of the processing stage unit 112 is also referred to as intermediate feature or intermediate feature data.
- the output of the processing stage unit 112-1 is referred to as intermediate feature data Y1.
- the output of the processing stage unit 112-2 is referred to as intermediate feature data Y2.
- the output of the processing stage unit 112-3 is referred to as intermediate feature data Y3.
- Each intermediate feature data corresponds to a type of feature data. In the example of FIG. 2, the data channel-divided from the intermediate feature data also corresponds to a type of feature data.
- FIG. 3 is a schematic block diagram showing a configuration example of the processing stage unit 112.
- the processing stage unit 112 includes a downsampling unit 121 and a processing block unit 122.
- the processing stage unit 112 includes N processing block units 122. These N processing block units 122 are connected in series, and further connected to the downsampling unit 121 in series. When distinguishing N processing block portions, reference numerals 122-1, ..., 122-N are added in order from the upstream side to the downstream side of the data flow. N may be an integer of 1 or more.
- the downsampling unit 121 receives input of pixel format data (data indicated by a sequence of pixel values) and reduces the image size (number of pixels) of the input data. Specifically, the input data to the downsampling unit 121 is preprocessed image data or pixel-format feature data (data in which channels are divided).
- the method and reduction rate at which the downsampling unit 121 reduces the image size is not limited to a specific one.
- the downsampling unit 121 may reduce the number of pixels to a quarter of the image by replacing each of the four pixels of 2 vertical ⁇ 2 horizontal with 1 pixel. In that case, the downsampling unit 121 may select the maximum value among the pixel values of the four pixels.
- the downsampling unit 121 may calculate the average of the pixel values of the four pixels and use it as the pixel value of the image after the size reduction.
- the number of output channels may be set to four times the number of input channels. Then, the downsampling unit 121 may allocate four pixels of 2 vertical ⁇ 2 horizontal to different channels.
- the number of input channels referred to here is the number of channels in the input data to the downsampling unit 121.
- the number of output channels is the number of channels in the output data from the downsampling unit 121.
- FIG. 4 is a schematic block diagram showing a configuration example of the processing block unit 122.
- the processing block unit 122 includes an affine channel conversion unit 131, a channel division unit 132, a convolution processing unit 133, a multiplication unit 134, an addition unit 135, and a channel coupling unit 136.
- the affine channel conversion unit 131 corresponds to the affine layer (Affine Layer) in the convolutional neural network.
- the affine layer is also referred to as a fully connected layer.
- the affine channel conversion unit 131 weights the input to the processing block unit 122. This weighting corresponds to the weighting for the input to the neuron model, which is commonly performed in neural networks.
- the affine channel conversion unit 131 may perform processing using a filter having a size of 1 ⁇ 1.
- the channel division unit 132 divides the output of the affine channel conversion unit 131 into data for each channel. For example, the channel division unit 132 divides each channel included in the output data of the affine channel conversion unit 131 into one of two groups, group A and group B. The channel division unit 132 outputs the channels distributed to the group A to the multiplication unit 134, and outputs the channels distributed to the group B to the convolution processing unit 133 and the channel coupling unit 136.
- the channel referred to here may be feature data of individual images.
- the channel division may be to distribute the feature data of each image into any of a plurality of groups.
- the output data of the affine channel conversion unit 131 may include the feature data of a plurality of images, and the feature data of each image may be treated as a channel.
- the channel division unit 132 may divide the feature data of each image into any of a plurality of groups by dividing the channel.
- the convolution processing unit 133 receives the input of the data of the group B (data distributed to the group B), and performs the convolution processing on the input data.
- the convolution processing unit 133 may perform a series of processing such as convolution processing and non-linear conversion on the input data.
- the convolution processing unit 133 may be configured by using a convolutional neural network.
- the convolution processing unit 133 distributes the processed data into two groups, group C and group D.
- the convolution processing unit 133 outputs the data distributed to the group C to the multiplication unit 134, and outputs the data distributed to the group D to the addition unit 135.
- the multiplication unit 134 receives the input of the data of the group A and the data of the group C, and performs the multiplication of the data of the group A and the data of the group C for each element.
- the data of group A and the data of group C have the same number of vertical elements and the same number of horizontal elements, and the multiplication unit 134 is used for each element at the same position between the data of group A and the data of group C. , Multiply the value of the element.
- the multiplication unit 134 outputs the data of the multiplication result to the addition unit 135.
- the addition unit 135 receives the input of the data from the multiplication unit 134 and the data of the group D, and adds the input data from the multiplication unit 134 and the data of the group D. Specifically, the addition unit 135 adds the data from the multiplication unit 134 and the data of the group D for each element.
- the data from the multiplication unit 134 and the data in the group D have the same number of vertical elements and the same number of horizontal elements, and the addition unit 135 has the same position of the data from the multiplication unit 134 and the data in the group D. For each element of, add the value of the element.
- the addition unit 135 outputs the data of the addition result to the channel coupling unit 136.
- the channel coupling unit 136 performs the reverse processing to the processing performed by the channel dividing unit 132. As a result, the channel coupling unit 136 combines one data from the adding unit 135 and one data of the group B into one data.
- the reverse processing referred to here is a processing corresponding to an inverse operation.
- the combination referred to here may be to combine a plurality of data into one data in a divisible manner.
- Each of the channel division units 113 of the feature extraction unit 12 allocates each of the intermediate features output by the processing stage unit 112 to one of the two groups. As a result, the channel division unit 113 extracts data to be collected in the feature data group as communication data to the receiving side device 20 from the intermediate feature data output by the processing stage section 112.
- the channel may be feature data of individual images.
- the channel division may be to distribute the feature data of each image into any of a plurality of groups.
- the communication data generation unit 13 generates communication data based on the feature data. Specifically, the communication data generation unit 13 converts the feature data group output by the feature extraction unit 12 into communication data.
- the communication data generation unit 13 corresponds to an example of a communication data generation means.
- the quantization unit 14 quantizes the feature data of the input image.
- the quantization referred to here may be rounding (rounding, rounding down, or rounding up) from a real number to an integer. Therefore, the quantization of the feature data performed by the quantization unit 14 is to convert each of the real numbers included in the feature data into an integer.
- the real number included in the feature data may be a further element of the real number vector which is an element of the feature data.
- the quantization unit 14 corresponds to an example of a quantization means.
- the coding unit 15 entropy-encodes the quantized feature data.
- the entropy coding referred to here is to perform data conversion (coding) so as to minimize the information entropy based on the predicted probability distribution of the input data (input code).
- a known entropy coding algorithm can be used for the processing performed by the coding unit 15.
- the coding unit 15 converts the feature data into a bit stream (Bit Stream, a data stream represented by a bit string) by entropy coding.
- Bit Stream a data stream represented by a bit string
- the coding method used by the information processing system 1 is not limited to the entropy coding method.
- Various coding methods capable of generating data suitable for communication such as a bit stream can be applied to the information processing system 1.
- Neither the quantization performed by the quantization unit 14 nor the coding performed by the coding unit 15 is limited to a specific process. By combining these processes, various processes that can convert feature data into a bit stream for transmission can be used.
- the transmission unit 16 transmits communication data. Specifically, the transmission unit 16 transmits the bit stream output by the coding unit 15 to the reception unit 21 of the reception side device 20 as a communication signal.
- the transmission unit 16 corresponds to an example of transmission means.
- the communication method between the transmitting unit 16 and the receiving unit 21 is not limited to a specific one.
- the transmitting unit 16 and the receiving unit 21 may perform wireless communication, or may perform wired communication.
- the receiving unit 21 receives communication data based on the feature data of the acquired image. Specifically, the receiving unit 21 receives the signal from the transmitting unit 16 and restores the bit stream.
- the receiving unit 21 corresponds to an example of receiving means.
- the feature restoration unit 22 restores the feature data based on the communication data received by the reception unit 21.
- the feature restoration unit 22 corresponds to an example of the feature restoration means.
- the decoding unit 23 converts the bitstream into quantized feature data by entropy decoding.
- the decoding performed by the decoding unit 23 corresponds to the inverse operation of the coding performed by the coding unit 15.
- the coding method used by the information processing system 1 is not limited to the entropy coding method.
- the decoding performed by the receiving side device 20 is not limited to the entropy decoding, and may be any one that decodes the data encoded by the transmitting side device 10.
- the dequantization unit 24 dequantizes the quantized feature data acquired by the decoding unit 23. Specifically, the dequantization unit 24 converts each of the integers included in the feature data into a real number.
- the method by which the dequantization unit 24 converts an integer into a real number is not limited to a specific method.
- the dequantization unit 24 may store in advance a probability distribution representing the coding probability of the real number vector as an element of the feature data, and perform sampling based on this probability distribution.
- the probability distribution representing the coding probability of the real vector as an element of the feature data corresponds to the example of the probability distribution of the feature data before quantization.
- the dequantization unit 24 can perform dequantization with high accuracy by reflecting the probability distribution of the feature data in the dequantization.
- the dequantization unit 24 may change only the data format from integer data to real data while leaving the integer value as it is.
- the dequantization unit 24 corresponds to an example of the dequantization means.
- the dequantization performed by the dequantization unit 24 is ideally the inverse operation of the quantization by the quantization unit 14, but usually, the value before quantization on the transmitting side is always accurately restored on the receiving side. Can't. It is considered that the feature data after dequantization by the dequantization unit 24 also includes quantization noise (quantization error). Quantization noise is an error due to quantization and dequantization. When indicating that it contains quantization noise, "Noisy" is added to the term, such as "Noisy feature data" and "Noisy intermediate feature data”.
- the quantization noise included in the noisy feature data for the restoration and image recognition of the received image performed by the receiving device 20 When the size of the real number included in the feature data is larger than the size of the rounding in the quantization, the quantization noise included in the noisy feature data for the restoration and image recognition of the received image performed by the receiving device 20. The effect of is small.
- the size of the real number included in the feature data may be increased according to the required accuracy.
- Increasing the size of the real number included in the feature data is performed, for example, by taking a large upper limit of the pixel value in the acquired image and expressing the pixel value with a large value.
- the dequantization by the dequantization unit 24 can be regarded as an approximate inverse operation with respect to the quantization by the quantization unit 14.
- the intermediate feature generation unit 25 calculates the noisy intermediate feature data from the noisy feature data group output by the dequantization unit 24.
- the operation of the intermediate feature generation unit 25 is ideally the inverse operation of the operation of the feature extraction unit 12, but is not limited to this.
- the intermediate feature generation unit 25 may be capable of calculating noisy intermediate feature data with the required accuracy according to the application of the information processing system 1.
- the intermediate feature generation unit 25 is configured by using a deep learning model by a convolutional neural network capable of inverse calculation, and the intermediate feature generation unit 25 is a channel division unit 113-1 and a processing stage unit of the feature extraction units 12.
- 112-2, the channel division unit 113-2, and the processing stage unit 112-3 are inverse models.
- the inverse model referred to here is a model that performs an inverse operation. That is, a case where the intermediate feature generation unit 25 performs an inverse operation on the operation by the above-mentioned part of the feature extraction unit 12 will be described as an example.
- FIG. 5 is a schematic block diagram showing a configuration example of the intermediate feature generation unit 25.
- the intermediate feature generation unit 25 includes a reverse processing stage unit 211 and a channel coupling unit 212.
- the two reverse processing stage portions 211 and the two channel coupling portions 212 are alternately arranged and connected in series.
- reference numerals 211-1 and 211-2 are attached in order from the upstream side to the downstream side of the data flow.
- reference numerals 212-1 and 212-2 are attached in order from the upstream side to the downstream side of the data flow.
- Each of the reverse processing stage units 211 performs the reverse calculation of the operation of one processing stage unit 112.
- the reverse processing stage unit 211-1 performs the reverse calculation of the calculation of the processing stage unit 112-3.
- the reverse processing stage unit 211-2 performs the reverse calculation of the calculation of the processing stage unit 112-2.
- the noisy feature data group input to the intermediate feature generation unit 25 includes the noisy intermediate feature data Y1'.
- the noisy intermediate feature data Y1' is data in which the intermediate feature data Y3 output by the processing stage unit 112-3 (FIG. 2) is restored including quantization noise.
- noisy intermediate feature data Y2' The output of the channel coupling portion 212-1 is referred to as noisy intermediate feature data Y2'.
- the noisy intermediate feature data Y2' is data in which the intermediate feature data Y2 output by the processing stage unit 112-2 is restored including quantization noise.
- the output of the channel coupling portion 212-2 is referred to as noisy intermediate feature data Y3'.
- the noisy intermediate feature data Y3' is data in which the intermediate feature data Y1 output by the processing stage unit 112-1 is restored including quantization noise.
- FIG. 6 is a schematic block diagram showing a configuration example of the reverse processing stage unit 211.
- the reverse processing stage unit 211 includes a reverse processing block unit 221 and an upsampling unit 222.
- N reverse processing block units 221 are connected in series, and an upsampling unit 222 is further connected in series.
- reference numerals 221-1, ..., 221-N are added in order from the upstream side to the downstream side of the data flow.
- Each of the reverse processing block units 221 performs the reverse calculation of the operation of one processing block unit 122.
- the reverse processing block units 221-1, ..., 221-N perform reverse operations of the operations of the processing block units 122-N, ..., 122-1, respectively.
- FIG. 7 is a schematic block diagram showing a configuration example of the reverse processing block unit 221.
- the reverse processing block unit 221 includes a channel division unit 231, a convolution processing unit 232, a subtraction unit 233, a division unit 234, a channel coupling unit 235, and an inverse affine channel conversion unit 236. Be prepared.
- the channel dividing unit 231 performs the inverse operation of the operation of the channel connecting unit 136. As a result, the channel division unit 231 performs the same processing as the channel division unit 132. For example, the channel division unit 231 divides each channel included in the input data to the channel division unit 231 itself into one of two groups, group A'and group B', like the channel division unit 132. Group A'is a group corresponding to group A. Group B'is a group corresponding to group B. The channel division unit 231 outputs the data distributed to the group A'to the subtraction unit 233, and outputs the data distributed to the group B'to the convolution processing unit 232 and the channel coupling unit 235.
- the convolution processing unit 232 performs the same processing as the convolution processing unit 133. Specifically, the convolution processing unit 232 receives the input of the data of the group B'and performs the convolution processing on the input data. When the convolution processing unit 133 performs a series of processing such as a convolution processing and a non-linear conversion on the input data, the convolution processing unit 232 also performs a series of processing similar to the convolution processing unit 133.
- the convolution processing unit 232 may be configured by using a convolutional neural network.
- the convolution processing unit 232 distributes the processed data into two groups, group C'and group D'.
- Group C' is a group corresponding to Group C.
- Group D' is a group corresponding to group D.
- the convolution processing unit 232 outputs the data distributed to the group D'to the subtraction unit 233, and outputs the data distributed to the group C'to the division unit 234.
- the subtraction unit 233 performs the inverse calculation of the addition unit 135. Specifically, the subtraction unit 233 receives the input of the data of the group A'and the data of the group D', and subtracts the data of the group D'from the input data of the group A'. More specifically, the subtraction unit 233 subtracts the value of the data element of the group D'from the value of the data element of the group A'for each element of the data of the group A'and the data of the group D'. ..
- the data of group A'and the data of group D' are the same in both the number of vertical elements and the number of horizontal elements, and the subtraction unit 233 is in the same position as the data of group A'and the data of group D'. For each element of, the value of the data element of group D'is subtracted from the value of the data element of group A'.
- the subtraction unit 233 outputs the data of the subtraction result to the division unit 234.
- the division unit 234 performs the inverse calculation of the multiplication unit 134. Specifically, the subtraction unit 234 receives the input of the data from the subtraction unit 233 and the data of the group C', and the subtraction unit 233 for each element of the data from the subtraction unit 233 and the data of the group C'. The value of the data element from is divided by the value of the data element of group C'. The data from the subtraction unit 233 and the data in the group C'are the same in both the number of vertical elements and the number of horizontal elements, and the division unit 234 is the data from the subtraction unit 233 and the data in the group C'. For each element at the same position, the value of the data element from the subtraction unit 233 is divided by the value of the data element of group C'. The division unit 234 outputs the division result data to the channel coupling unit 235.
- the channel coupling unit 235 performs the reverse processing with respect to the processing performed by the channel dividing unit 231. As a result, the channel coupling unit 235 combines one data from the division unit 234 and one data of the group B'to one data. The processing of the channel coupling unit 235 also corresponds to the reverse processing with respect to the processing performed by the channel dividing unit 132.
- the inverse affine channel conversion unit 236 performs the inverse operation of the operation of the affine channel conversion unit 131.
- the upsampling unit 222 of the reverse processing stage unit 211 ideally performs the reverse operation of the operation of the downsampling unit 121.
- the data before downsampling on the transmitting side cannot always be accurately restored on the receiving side.
- the downsampling unit 121 replaces four pixels with one pixel having an average pixel value of the pixel values of those four pixels.
- the upsampling unit 222 cannot usually calculate the original four pixel values from the obtained one pixel value.
- the upsampling unit 222 may approximately restore the data before downsampling. For example, the upsampling unit 222 divides each pixel of the input data into four pixels of 2 vertical ⁇ 2 horizontal, and sets the value of each pixel to the same value as the value of the original pixel to obtain the data ( Image data or feature data) may be converted into image data having a size of four times.
- the channel coupling unit 212 of the intermediate feature generation unit 25 performs an inverse operation with respect to the operation of the channel dividing unit 113. As a result, the channel coupling unit 212 generates data in which a plurality of channels are combined into one.
- the channel coupling unit 212-1 performs an inverse operation on the operation of the channel dividing unit 113-2.
- the channel coupling unit 212-2 performs an inverse operation on the operation of the channel dividing unit 113-1.
- the acquired image restoration unit 26 calculates an image based on the intermediate feature data output by the intermediate feature generation unit 25. Specifically, the acquired image restoration unit 26 restores the acquired image by performing the reverse processing of the processing of the preprocessing unit 111 and the processing stage unit 112-1 among the processing of the feature extraction unit 12.
- the image calculated by the acquired image restoration unit 26 is also referred to as a restored image.
- the acquired image restoration unit 26 corresponds to an example of the target restoration means. Restoration of the acquired image by the acquired image restoration unit 26 corresponds to a process of restoring the acquired image data based on the feature data restored by the feature restoration unit 22.
- FIG. 8 is a schematic block diagram showing a configuration example of the acquired image restoration unit 26.
- the acquired image restoration unit 26 includes a reverse processing stage unit 211 and a post-processing unit 241.
- the reverse processing stage unit 211 of the acquired image restoration unit 26 is distinguished from the reverse processing stage unit (FIG. 5) of the intermediate feature generation unit 25, the reverse processing stage unit 211 of the acquired image restoration unit 26 is the reverse processing stage unit 211. Notated as -3.
- the reverse processing stage unit 211-3 corresponds to the reverse model of the processing stage unit 112-1 (FIG. 2).
- the post-processing unit 241 performs an inverse operation of the operation of the pre-processing unit 111.
- the restored image is similar to the acquired image. Specifically, the restored image is an image in which quantization noise is added to the acquired image.
- the recognition unit 27 performs image recognition based on the noisy intermediate feature data group output by the intermediate feature generation unit 25.
- the noisy intermediate feature data group output by the intermediate feature generation unit 25 corresponds to the feature data of the restored image.
- the image recognition performed by the recognition unit 27 corresponds to image recognition for the restored image.
- Image recognition for the restored image can be said to be image recognition for the acquired image which is the original image of the restored image. Therefore, the image recognition performed by the recognition unit 27 corresponds to performing the recognition process for the acquired image, which is the expression content of the acquired image data, based on the feature data restored by the feature restoration unit 22.
- the recognition unit 27 corresponds to an example of the recognition means.
- FIG. 9 is a schematic block diagram showing a configuration example of the recognition unit 27.
- the recognition unit 27 includes an intermediate feature processing unit 251, an upsampling unit 252, an addition unit 253, a position estimation processing unit 254, a classification processing unit 255, and an NMS (Non-Maximum Suppression). It is provided with a processing unit 256.
- one position estimation processing unit 254 and one classification processing unit 255 are connected to each of the three intermediate feature processing units 251.
- the output of the first intermediate feature processing unit 251 is input to the first upsampling unit 252, and the output of the upsampling unit 252 and the output of the second intermediate feature processing unit 251 are set to 1.
- the second addition unit 253 adds each pixel.
- the data after addition is input to the second upsampling unit 252, and the output of the upsampling unit 252 and the output of the third intermediate feature processing unit 251 are pixelated by the second addition unit 253. It is added every time.
- the first intermediate feature processing unit 251 described above is referred to as an intermediate feature processing unit 251-1.
- the second intermediate feature processing unit 251 is referred to as an intermediate feature processing unit 251-2.
- the third intermediate feature processing unit 251 is referred to as an intermediate feature processing unit 251-3.
- the position estimation processing unit 254 connected to the intermediate feature processing unit 251-1 is referred to as the position estimation processing unit 254-1.
- the position estimation processing unit 254 connected to the intermediate feature processing unit 251-2 is referred to as a position estimation processing unit 254-2.
- the position estimation processing unit 254 connected to the intermediate feature processing unit 251-3 is referred to as a position estimation processing unit 254-3.
- the classification processing unit 255 connected to the intermediate feature processing unit 251-1 is referred to as the classification processing unit 255-1.
- the classification processing unit 255 connected to the intermediate feature processing unit 251-2 is referred to as a classification processing unit 255-2.
- the classification processing unit 255 connected to the intermediate feature processing unit 251-3 is referred to as a classification processing unit 255-3.
- the upsampling unit 252 to which the output of the intermediate feature processing unit 251-1 is input is referred to as the upsampling unit 252-1.
- the upsampling unit 252 to which the output of the intermediate feature processing unit 251-2 is input is referred to as an upsampling unit 252-2.
- addition unit 253 that adds the output of the intermediate feature processing unit 251-2 and the output of the upsampling unit 252-1 is referred to as an addition unit 253-1.
- addition unit 253 that adds the output of the intermediate feature processing unit 251-3 and the output of the upsampling unit 252-2 is referred to as an addition unit 253-2.
- Each of the intermediate feature processing units 251 detects a recognition target in the noisy intermediate feature included in the noisy intermediate feature data.
- the intermediate feature processing unit 251 may not detect any recognition target. Further, one intermediate feature processing unit 251 may detect a plurality of recognition targets. A known method can be used as a method for the intermediate feature processing unit 251 to detect the recognition target.
- Each of the upsampling units 252 performs the same processing as the upsampling unit 222 (FIG. 6) of the reverse processing stage unit 211.
- the upsampling unit 252 restores the data before downsampling by the downsampling unit 121, as in the case of the upsampling unit 222.
- the upsampling unit 252 may approximately restore the data before downsampling by the downsampling unit 121.
- Each of the addition units 253 adds the output of the intermediate feature processing unit 251 and the output of the upsampling unit 252 for each pixel.
- Each of the position estimation processing units 254 estimates the position in the restored image of the recognition target detected by the intermediate feature processing unit 251.
- a known method can be used as a method for the position estimation processing unit 254 to detect the position in the restored image to be recognized.
- the classification processing unit 255 classifies the recognition target detected by the intermediate feature processing unit 251. This classification may be an estimation of the type of recognition target. A known method can be used as a method for the classification processing unit 255 to classify the recognition target.
- the NMS processing unit 256 When the regions recognized as the same class overlap on the image (here, on the restored image), the NMS processing unit 256 eliminates the overlap.
- the NMS processing unit 256 may delete any one of the overlapping regions of the same class and delete the other.
- the NMS processing unit 256 may replace the overlapping regions with one region that includes those regions.
- a method for the NMS processing unit 256 to perform the processing a method known as Non-Maximum Suppression may be used.
- the output unit 28 outputs information indicating the restored image generated by the acquired image restoration unit 26 and the recognition result by the recognition unit 27.
- the output unit 28 may include a display device to display the restored image. Then, the output unit 28 may indicate the recognition target in the restored image by surrounding it with a bounding box (a rectangle that just surrounds the area), and indicate the recognition target class by the color of the bounding box.
- the method in which the output unit 28 outputs the restored image and the recognition result is not limited to a specific method.
- the output unit 28 may output the restored image and the recognition result separately.
- the output unit 28 corresponds to an example of output means.
- FIG. 10 is a flowchart showing an example of a processing procedure performed by the transmitting side device 10.
- the transmitting device 10 may repeat the process of FIG. For example, when the transmitting side device 10 repeats the acquisition of the still image at a predetermined cycle, the process of FIG. 10 may be performed every time the still image is acquired.
- the image acquisition unit 11 acquires an image (step S101). As described above, the image acquired by the image acquisition unit 11 is also referred to as an acquired image.
- the feature extraction unit 12 extracts the feature data of the acquired image (step S102).
- the quantization unit 14 quantizes the feature data (step S103).
- the coding unit 15 encodes the quantized feature data (step S104).
- the coding unit 15 converts the quantized feature data into a bit stream by encoding the quantized feature data.
- the transmission unit 16 transmits the bit stream output by the coding unit 15 to the receiving device 20 (step S105). After step S105, the transmitting device 10 ends the process of FIG.
- FIG. 11 is a flowchart showing an example of a processing procedure performed by the receiving side device 20.
- the receiving side device 20 may repeat the processing of FIG. 11 in response to the repetition of the processing of FIG. 10 by the transmitting side device 10.
- the receiving unit 21 receives the bit stream (step S201).
- the decoding unit 23 decodes the bit stream received by the receiving unit 21 (step S202). As described above, the decoding unit 23 decodes by the inverse operation of the encoding performed by the coding unit 15 of the transmitting side device 10. The decoding unit 23 generates quantized feature data by decoding the bit stream.
- the dequantization unit 24 calculates the noisy feature data by dequantizing the data obtained by decoding the bitstream in step S202 (step S203). As described above, it can be said that the noisy feature data is obtained by adding quantization noise to the feature data extracted by the feature extraction unit 12.
- the intermediate feature generation unit 25 generates noisy intermediate feature data based on the noisy feature data (step S204).
- the acquired image restoration unit 26 generates a restored image based on the noisy intermediate feature data (step S205).
- the recognition unit 27 performs image recognition based on the noisy intermediate feature data and calculates the recognition result (step S206).
- the output unit 28 outputs the restored image and the recognition result (step S207).
- the receiving device 20 ends the process of FIG.
- the receiving unit 21 receives the communication data based on the feature data indicating the features of the acquired image, which is the expression content of the acquired image data.
- the feature restoration unit 22 restores the feature data based on the received communication data.
- the acquired image restoration unit 26 restores the acquired image data based on the restored data.
- the recognition unit 27 performs image recognition on the acquired image, which is the expression content of the acquired image data, based on the restored feature data.
- the output unit 28 outputs information indicating the expression content of the restored target data and the recognition result by the recognition process.
- the receiving side device 20 uses the feature data restored by the feature restoring unit 22 for both the restoration of the acquired image by the acquired image restoration unit 26 and the image recognition by the recognition unit 27. According to the receiving device 20, in comparison with the case where the image is restored and then the image recognition is performed using the restored image, the acquired image data is restored and the restored data is expressed. The processing time for performing image recognition on an image can be shortened.
- the receiving unit 21 receives communication data based on the quantized feature data.
- the dequantization unit 24 dequantizes the quantized feature data based on sampling according to the probability distribution of the feature data before it is quantized. It is expected that the dequantization unit 24 can perform dequantization with high accuracy by reflecting the probability distribution of the feature data in the dequantization.
- the receiving unit 21 receives communication data based on the intermediate feature data Y1 and the intermediate feature data Y2 calculated based on the data downsampled by the downsampling unit 121 from the intermediate feature data Y1.
- the feature restoration unit 22 restores the noisy intermediate feature data Y3'based on the data upsampled by the upsampling unit 222 from the noisy intermediate feature data Y2'restored based on the communication data received by the intermediate feature data Y2. ..
- the receiving side device 20 restores the acquired image data using the feature data having different image sizes, so that the image compression ratio of the transmitting side device 10 can be adjusted relatively easily.
- the feature restoration unit 22 uses a process corresponding to the inverse calculation of the process in which the processing stage unit 112 calculates the intermediate feature data Y2 based on the data downsampled from the intermediate feature data Y1. To restore. As a result, it is expected that the feature restoration unit 22 can restore the intermediate feature data with relatively high accuracy.
- FIG. 12 is a schematic block diagram showing a configuration example of the information processing system according to the second embodiment.
- the information processing system 2 includes a transmitting side device 30 and a receiving side device 40.
- the transmission side device 30 includes an image acquisition unit 11, a feature extraction unit 12, a communication data generation unit 31, a transmission unit 16, and a noisy feature data storage unit 35.
- the communication data generation unit 31 includes a quantization unit 14, a coding unit 15, a dequantization unit 32, a feature difference calculation unit 33, and a feature calculation unit 34.
- the receiving device 20 includes a receiving unit 21, a feature restoration unit 41, an acquired image restoration unit 26, a recognition unit 27, an output unit 28, and a noisy feature data storage unit 43.
- the feature restoration unit 41 includes a decoding unit 23, a dequantization unit 24, an intermediate feature generation unit 25, and a feature calculation unit 42.
- the parts having the same functions corresponding to the parts of FIG. 1 have the same reference numerals (11, 12, 14, 15, 16, 21, 23, 24, 25, 26, 27, 28). Is added, and detailed description thereof will be omitted here. Comparing the configuration of the information processing system 2 shown in FIG. 12 with the information processing system 1 shown in FIG. 1, a functional unit for efficiently transmitting and processing a moving image is added. In other respects, the information processing system 2 is similar to the information processing system 1.
- the image acquisition unit 11 acquires a moving image or a still image that is repeatedly captured in a relatively short cycle such as a 1-second cycle.
- the image acquisition unit 11 acquires a moving image
- the data of each frame of the moving image is treated as the acquired image data.
- One of the acquired image data is referred to as the first acquired image data
- the data of the acquired image captured after the first acquired image is referred to as the second acquired image data.
- the first acquired image data corresponds to the example of the first target data.
- the second acquired image data corresponds to the example of the second target data.
- the feature extraction unit 12 calculates the feature data of each of the plurality of images acquired by the image acquisition unit 11 (when the image acquisition unit 11 acquires a moving image, the frame of the moving image). For example, the feature extraction unit 12 extracts the first feature data from the first acquired image data and extracts the second feature data from the second acquired image data.
- the communication data generation unit 31 converts the feature data (for example, the feature data group) of the image into the communication data, as in the communication data generation unit 13 of the first embodiment. ..
- the communication data generation unit 31 calculates the feature difference data for the second and subsequent images acquired by the image acquisition unit 11, and generates the communication data based on the calculated feature difference data.
- the feature difference data is data indicating the difference between the two feature data calculated by the feature extraction unit 12.
- the communication data generation unit 31 calculates the feature difference data indicating the difference between the first feature data and the second feature data, and generates the communication data based on the calculated feature difference data.
- the communication data generation unit 31 generates noisy feature difference data including quantization noise by the quantization in the quantization unit 14 and the dequantization in the dequantization unit 32, and is based on the noisy feature difference data. Generate communication data.
- the dequantization unit 32 performs the same processing as the dequantization unit 24 of the receiving side device 40. As a result, the dequantization unit 32 generates the same noisy feature data as the noisy feature data generated by the dequantization unit 24.
- the noisy feature data storage unit 35 temporarily stores the noisy feature data.
- the noisy feature data stored in the noisy feature data storage unit 35 is used to generate noisy feature difference data in the next process.
- the next processing referred to here is processing for the next image among the processing for each image acquired by the image acquisition unit 11, such as processing for each frame of the moving image acquired by the image acquisition unit 11.
- the feature difference calculation unit 33 calculates the noisy feature difference data.
- the noisy feature difference data is the difference data between the feature data generated in the continuous processing and the noisy feature data generated in the previous processing.
- the transmitting side device 30 transmits a bit stream obtained by quantizing and encoding noisy feature difference data instead of the feature data to the receiving side device 40.
- the receiving device 40 restores the noisy feature difference data from the received bitstream.
- the receiving side device 40 calculates the noisy feature data in the current process by adding the restored noisy feature difference data and the noisy feature data in the previous process. Subsequent processing is the same as in the case of the receiving side device 20 of the first embodiment.
- the receiving device 40 corresponds to an example of an information processing device.
- the feature calculation unit 34 of the transmission side device 30 stores the noisy feature difference data calculated by the dequantization unit 32 in this processing and the noisy feature data storage unit 35.
- the noisy feature data in the current process is calculated by adding the noisy feature data in the previous process.
- the feature calculation unit 34 updates the noisy feature data stored in the noisy feature data storage unit 35 in the previous process to the noisy feature data in the current process calculated by the feature calculation unit 34 itself.
- the update of the data referred to here may be overwriting of the data.
- the noisy feature data storage unit 43 of the receiving side device 40 temporarily stores the noisy feature data, like the noisy feature data storage unit 35 of the transmitting side device 30.
- the feature calculation unit 42 adds the noisy feature difference data in the current process restored by the dequantization unit 24 and the noisy feature data in the previous process stored in the noisy feature data storage unit 43. As a result, the feature calculation unit 42 calculates the noisy feature data in this process.
- the feature calculation unit 42 outputs the calculated noisy feature data to the intermediate feature generation unit 25. Further, the feature calculation unit 42 updates the noisy feature data stored in the noisy feature data storage unit 43 in the previous process to the noisy feature data in the current process calculated by the feature calculation unit 42 itself.
- FIG. 13 is a schematic block diagram showing a configuration example of the feature difference calculation unit 33.
- the feature difference calculation unit 33 includes a difference processing stage unit 311 and an upsampling unit 312.
- FIG. 13 shows an example in which the feature difference calculation unit 33 is configured by using an inverted deep convolutional neural network model.
- the configuration of the feature difference calculation unit 33 is not limited to a specific one.
- the feature difference calculation unit 33 includes three difference processing stage units 311 and two upsampling units 312. These are connected in series in an arrangement in which one upsampling unit 312 is provided between each of the two difference processing stage units 311.
- reference numerals 311-1, 311-2, and 311-3 are attached in order from the upstream side to the downstream side of the data flow.
- reference numerals 312-1 and 312-2 are attached in order from the upstream side to the downstream side of the data flow.
- each of the difference processing stage units 311 calculates the difference between the feature data in the time step t and the noisy feature data in the time step t-1.
- FIG. 14 is a schematic block diagram showing a configuration example of the difference processing stage unit 311.
- the difference processing stage unit 311 includes a difference processing block unit 321.
- N difference processing block units 321 are connected in series.
- reference numerals 321-1, ..., 321-N are added in order from the upstream side to the downstream side of the data flow.
- FIG. 15 is a schematic block diagram showing a configuration example of the difference processing block unit 321.
- the difference processing block unit 321 includes an affine channel conversion unit 331, a channel division unit 332, a convolution processing unit 333, a multiplication unit 334, an addition unit 335, and a channel coupling unit 336. ..
- the affine channel conversion unit 331, the channel division unit 332, the multiplication unit 334, the addition unit 335, and the channel coupling unit 336 are the affine channel conversion unit 131, the channel division unit 132, the multiplication unit 134, the addition unit 135, and , The same as the channel coupling portion 136.
- the affine channel conversion unit 331 performs the same processing as the affine channel conversion unit 131 on the data from the other difference processing block unit 321 or the feature data from the feature extraction unit 12.
- the convolution processing unit 333 receives input from the channel division unit 332, the noisy feature data in the time step t-1, and the data from the upsampling unit 312.
- the data from the channel dividing unit 332 to the convolution processing unit 333 is the data of the group corresponding to the group B. Further, the convolution processing unit 333 acquires the noisy feature data in the time step t-1 stored by the noisy feature data storage unit 35.
- the convolution processing unit 333 combines the data from the channel division unit 332, the noisy feature data in the time step t-1, and the data from the upsampling unit 312, and the convolution processing unit 133 refers to the combined data. Perform the same processing as. Specifically, the convolution processing unit 333 performs a convolution process on the combined data.
- the convolution processing unit 333 may perform a series of processing such as convolution processing and non-linear conversion on the combined data.
- the convolution processing unit 333 may be configured by using a convolutional neural network.
- the convolution processing unit 333 may combine the data from the channel division unit 332 with the noisy feature data in the time step t-1. ..
- the convolution processing unit 333 distributes the processed data into two groups, a group corresponding to group C and a group corresponding to group D.
- the convolution processing unit 333 outputs the data distributed to the group corresponding to the group C to the multiplication unit 334, and outputs the data distributed to the group corresponding to the group D to the addition unit 335.
- FIG. 16 is a schematic block diagram showing a configuration example of the feature calculation unit 34.
- the feature calculation unit 34 includes a restoration processing stage unit 341 and an upsampling unit 342.
- FIG. 16 shows an example in which the feature calculation unit 34 is configured by using an inverted deep convolutional neural network model.
- the configuration of the feature calculation unit 34 is not limited to a specific one.
- the feature calculation unit 34 includes three restoration processing stage units 341 and two upsampling units 342. These are connected in series in an arrangement in which one upsampling unit 342 is provided between each of the two restoration processing stage units 341.
- reference numerals 341-1, 341-2, and 341-3 are attached in this order from the upstream side to the downstream side of the data flow.
- reference numerals 342-1 and 342-2 are attached in order from the upstream side to the downstream side of the data flow.
- Each of the restoration processing stage units 341 calculates the noisy feature data in the time step t based on the feature data in the time step t-1 and the noisy feature difference data in the time step t.
- FIG. 17 is a schematic block diagram showing a configuration example of the restoration processing stage unit 341.
- the restoration processing stage unit 341 includes a restoration processing block unit 351.
- N restoration processing block units 351 are connected in series.
- reference numerals 351-1, ..., 351-N are added in order from the upstream side to the downstream side of the data flow.
- FIG. 18 is a schematic block diagram showing a configuration example of the restoration processing block unit 351.
- the restoration processing block unit 351 includes a channel division unit 361, a convolution processing unit 362, a subtraction unit 363, a division unit 364, a channel coupling unit 365, and an inverse affine channel conversion unit 366. Be prepared.
- the channel division unit 361, the subtraction unit 363, the division unit 364, the channel coupling unit 365, and the inverse affine channel conversion unit 366 are the channel division unit 231, the subtraction unit 233, the division unit 234, and the channel connection unit of the reverse processing block unit 221. This is the same as 235 and the inverse affine channel conversion unit 236.
- the channel division unit 361 performs the same processing as the channel division unit 231 on the data from the other restoration processing block unit 351 or the noisy feature difference data output by the dequantization unit 24.
- the process performed by the channel dividing unit 361 corresponds to the reverse processing of the process performed by the channel connecting unit 336.
- the operation performed by the subtraction unit 363 corresponds to an inverse operation with respect to the operation performed by the addition unit 335.
- the operation performed by the division unit 364 corresponds to an inverse operation with respect to the operation performed by the multiplication unit 334.
- the process performed by the channel coupling unit 365 corresponds to the reverse processing of the process performed by the channel dividing unit 332.
- the convolution processing unit 362 performs the same processing as the convolution processing unit 333. Specifically, the convolution processing unit 362 receives input from the channel division unit 361, the noisy feature data in the time step t-1, and the data from the upsampling unit 342. The data from the channel dividing unit 361 to the convolution processing unit 362 is the data of the group corresponding to the group B. Further, the convolution processing unit 362 acquires the noisy feature data in the time step t-1 stored by the noisy feature data storage unit 35.
- the convolution processing unit 362 combines the data from the channel division unit 361, the noisy feature data in the time step t-1, and the data from the upsampling unit 342, and the convolution processing unit 333 refers to the combined data. Perform the same processing as. Specifically, the convolution processing unit 362 performs a convolution process on the combined data.
- the convolution processing unit 362 may perform a series of processing such as convolution processing and non-linear conversion on the combined data.
- the convolution processing unit 362 may be configured by using a convolutional neural network.
- the convolution processing unit 362 divides the processed data into two groups, a group corresponding to group C and a group corresponding to group D.
- the convolution processing unit 362 outputs the data distributed to the group corresponding to the group D to the subtraction unit 363, and outputs the data distributed to the group corresponding to the group C to the division unit 364.
- the feature restoration unit 41 of the receiving device 40 restores the feature difference data based on the communication data received by the receiving unit 21, and the restored feature difference data and the noisy feature data storage unit 43 store the time step t-. Based on the noisy feature data in 1, the feature data in time step t is restored.
- the feature restoration unit 41 corresponds to an example of the feature restoration means.
- the transmitting side device 30 and the receiving side device 40 transmit and receive communication data indicating the feature difference data, so that the dequantization unit 24 dequantizes the quantized feature difference data. ..
- the dequantization unit 24 performs dequantization based on sampling according to the probability distribution of the feature difference data before being quantized. You may do it.
- the dequantization unit 24 may store in advance a probability distribution representing the coding probability of the real number vector as an element of the feature difference data, and perform sampling based on this probability distribution.
- the feature calculation unit 42 of the receiving side device 40 is the same as the feature calculation unit 34 of the transmitting side device 30.
- the transmitting side device 30 and the receiving side device 40 perform the same processing to generate and store the noisy feature data.
- the transmitting side device 30 uses the noisy feature data stored in the noisy feature data storage unit 35 as the previous noisy feature data (time step t-1) in calculating the noisy feature difference data.
- the receiving device 40 uses the previous noisy feature data (time step t-1) stored in the noisy feature data storage unit 43 when restoring the noisy feature data (time step t) from the noisy feature difference data. .. It is expected that the receiving side device 40 can restore the current noisy feature data with high accuracy by restoring the current noisy feature data using the previous noisy feature data similar to the transmitting side device 30.
- FIG. 19 is a flowchart showing an example of a processing procedure performed by the transmitting side device 30.
- FIG. 19 shows a procedure for processing one image when the transmitting side device 30 transmits a plurality of images (frames in the case of a moving image) such as a moving image or a continuously shot still image to the receiving side device 40. An example of is shown.
- the transmitting side device 30 repeats the process of FIG. 19 for each image.
- the image acquisition unit 11 acquires an image (step S301).
- the image acquired by the image acquisition unit 11 is also referred to as an acquired image.
- the time step of this processing is defined as the time step t. t is a positive integer.
- the feature extraction unit 12 extracts the feature data of the acquired image (step S302).
- the quantization unit 14 quantizes the feature data (step S311).
- the coding unit 15 encodes the quantized data (step S331).
- the “quantized data” is the difference data quantized in step S322.
- the coding unit 15 generates a bit stream for transmission by encoding the quantized data.
- the transmission unit 16 transmits the bit stream generated by the coding unit 15 to the receiving device 40 (step S332).
- step S333 YES
- the dequantization unit 32 calculates the noisy feature data by dequantizing the quantized data, and stores the noisy feature data storage unit 35. It is memorized (step S341).
- step S341 the quantization unit 14 has quantized the feature data in step S311. From this, noisy feature data can be obtained by dequantization in step S341.
- step S341 the transmitting device 30 ends the process of FIG.
- the feature difference calculation unit 33 calculates the feature difference data (step S321). Specifically, the feature difference calculation unit 33 reads out the noisy feature data stored in the noisy feature data storage unit 35. Since this noisy feature data was obtained in the previous execution of the process of FIG. 19 by the transmitting side device 30, it is the noisy feature data of time step t-1.
- the feature difference calculation unit 33 is based on the feature data (time step t) extracted by the feature extraction unit 12 in step S302 and the noisy feature data (time step t-1) read from the noisy feature data storage unit 35. To calculate the feature difference data.
- the quantization unit 14 quantizes the feature difference data (step S322). After step S322, the process proceeds to step S331.
- step S333 when the transmitting side device 30 determines in step S333 that t ⁇ 2 (step S333: NO), the dequantization unit 32 dequantizes the quantized data to obtain noisy feature difference data. Is calculated (step S351). When t ⁇ 2, the quantization unit 14 has quantized the feature difference data in step S322. From this, noisy feature difference data can be obtained by dequantization in step S351.
- the feature calculation unit 34 calculates the noisy feature data and stores it in the noisy feature data storage unit 35 (step S352). Specifically, the feature calculation unit 34 reads out the noisy feature data (time step t-1) stored in the noisy feature data storage unit 35.
- the feature calculation unit 34 includes the noisy feature difference data (time step t) calculated by the dequantization unit 32 in step S351 and the noisy feature data (time step t-1) read from the noisy feature data storage unit 35.
- the noisy feature data (time step t) is calculated based on.
- the feature calculation unit 34 stores the calculated noisy feature data (time step t) in the noisy feature data storage unit 35.
- FIG. 20 is a flowchart showing an example of a processing procedure performed by the receiving device 40.
- the receiving side device 40 repeats the processing of FIG. 20 in response to the repetition of the processing of FIG. 19 by the transmitting side device 30.
- Steps S401 and S402 of FIG. 20 are the same as steps S201 and S202 of FIG. 11 except that the bit stream may represent feature data and feature difference data.
- t ⁇ 2 quantized feature difference data can be obtained.
- step S403 the dequantization unit 24 calculates the noisy feature data and stores it in the noisy feature data storage unit 43 (step S411). Specifically, the dequantization unit 24 calculates the noisy feature data by dequantizing the data obtained by decoding the bit stream, as in the case of step S203 of FIG. Then, the dequantization unit 24 stores the calculated noisy feature data in the noisy feature data storage unit 43.
- step S411 the process proceeds to step S431. Steps S431 to S434 are the same as steps S204 to S207 of FIG. After step S434, the receiving device 40 ends the process of FIG.
- step S403 when the receiving device 40 determines in step S403 that t ⁇ 2 (step S403: NO), the dequantization unit 24 dequantizes the data obtained by decoding the bit stream in step S402. The noisy feature difference data is calculated (step S421).
- the feature calculation unit 42 calculates the noisy feature data (time step t) and stores it in the noisy feature data storage unit 43 (step S422). Specifically, the feature calculation unit 42 reads out the noisy feature data stored in the noisy feature data storage unit 43. Since this noisy feature data was obtained in the previous execution of the process of FIG. 20 by the receiving device 40, it is the noisy feature data of time step t-1.
- the feature calculation unit 42 includes the noisy feature difference data (time step t) calculated by the dequantization unit 24 in step S421 and the noisy feature data (time step t-1) read from the noisy feature data storage unit 35.
- the noisy feature data (time step t) is calculated based on.
- the feature calculation unit 42 stores the calculated noisy feature data in the noisy feature data storage unit 43. After step S422, the process proceeds to step S431.
- the receiving unit 21 shows the first feature data showing the features of the acquired image in the first time step, and the first feature data showing the features of the acquired image in the second time step, which is a time step later than the first time step.
- (Ii) Receive communication data based on the feature difference data indicating the difference from the feature data.
- the feature restoration unit 41 restores the feature difference data based on the received communication data, and restores the second feature data based on the restored feature difference data and the first feature data. According to the receiving device 40, by receiving the communication data based on the feature difference data, it is expected that the amount of communication can be reduced as compared with the case of receiving the communication data based on the feature data.
- the receiving unit 21 receives communication data based on the quantized difference data.
- the dequantization unit 24 dequantizes the quantized feature difference data based on sampling according to the probability distribution of the feature difference data before being quantized. It is expected that the dequantization unit 24 can perform dequantization with high accuracy by reflecting the probability distribution of the feature difference data in the dequantization.
- the processing settings performed by the transmitting device may be dynamically updated, such as dynamically changing the compression rate of communication data.
- the setting of the process performed by the receiving device may be dynamically updated.
- FIG. 21 is a schematic block diagram showing a first example of the configuration of the information processing system according to the third embodiment.
- the information processing system 3a includes a transmitting side device 51, a receiving side device 52, and a setting updating device 53.
- the setting update device 53 includes a setting update unit 54.
- the transmitting side device 51 and the receiving side device 52 may be the transmitting side device 10 and the receiving side device 20. That is, the third embodiment may be implemented based on the first embodiment. Alternatively, the transmitting side device 51 and the receiving side device 52 may be the transmitting side device 30 and the receiving side device 40. That is, the third embodiment may be implemented based on the second embodiment.
- the setting update unit 54 updates the processing setting of the transmitting side device 51 and the processing setting of the receiving side device 52.
- the setting update unit 54 dynamically updates the settings of these processes so that the processes of the feature extraction unit 12 and the processes of the intermediate feature generation unit 25 and the acquired image restoration unit 26 have an inverse operation relationship. do.
- the total number of the number of processing stage units 112 of the feature extraction unit 12 and the number of reverse processing stage units 211 of the intermediate feature generation unit 25 and the acquired image restoration unit 26 are the same.
- these numbers may be dynamically changed.
- the setting update unit 54 corresponds to an example of the setting update means. As a result, it is expected that the processing settings can be dynamically changed, such as dynamically changing the compression rate of the communication data, and that the receiving side device 52 can restore the feature data with high accuracy.
- the setting update unit 54 may be provided on either the transmitting side device or the receiving side device.
- FIG. 22 is a schematic block diagram showing a second example of the configuration of the information processing system according to the third embodiment. In the configuration shown in FIG. 22, in the information processing system 3b, the setting update unit 54 is provided in the transmitting side device 51. Other than that, the information processing system 3b is the same as that of the information processing system 3a.
- FIG. 23 is a schematic block diagram showing a third example of the configuration of the information processing system according to the third embodiment.
- the setting update unit 54 is provided in the receiving side device 52.
- the information processing system 3c is the same as that of the information processing system 3a.
- FIG. 24 is a schematic block diagram showing a configuration example of the information processing apparatus according to the fourth embodiment.
- the information processing device 610 includes a receiving unit 611, a feature restoration unit 612, an object restoration unit 613, a recognition unit 614, and an output unit 615.
- the receiving unit 611 receives communication data based on the feature data indicating the features of the expression content of the target data.
- the feature restoration unit 612 restores the feature data based on the received communication data.
- the target restoration unit 613 restores the target data based on the restored feature data.
- the recognition unit 614 performs recognition processing on the expression content of the target data based on the restored feature data.
- the output unit 615 outputs information indicating the expression content of the restored target data and the recognition result by the recognition process.
- the receiving unit 611 corresponds to the example of the receiving means
- the feature restoring unit 612 corresponds to the example of the feature restoring means.
- the target restoration unit 613 corresponds to an example of the target restoration means.
- the recognition unit 614 corresponds to an example of the recognition means.
- the output unit 615 corresponds to an example of output means.
- the information processing apparatus 610 uses the feature data restored by the feature restoration unit 612 for both the restoration of the target data by the target restoration unit 613 and the recognition process by the recognition unit 614. According to the information processing device 610, the restoration process of the target data and the expression content of the restored target data are compared with the case where the recognition process is performed using the restored target data after the target data is restored. The processing time for performing the recognition process is short.
- FIG. 25 is a schematic block diagram showing a configuration example of the information processing system according to the fifth embodiment.
- the information processing system 620 includes a transmitting side device 630 and a receiving side device 640.
- the transmission side device 630 includes a data acquisition unit 631, a feature extraction unit 632, a communication data generation unit 633, and a transmission unit 634.
- the receiving device 640 includes a receiving unit 641, a feature restoration unit 642, an object restoration unit 643, a recognition unit 644, and an output unit 645.
- the data acquisition unit 631 acquires the target data.
- the feature extraction unit 632 calculates feature data indicating the features of the expression content of the target data.
- the communication data generation unit 633 generates communication data based on the feature data.
- the transmission unit 634 transmits communication data.
- the receiving unit 641 receives the communication data.
- the feature restoration unit 642 restores the feature data based on the received communication data.
- the target restoration unit 643 restores the target data based on the restored feature data.
- the recognition unit 644 performs recognition processing for the expression content of the target data based on the restored feature data.
- the output unit 645 outputs information indicating the expression content of the restored target data and the recognition result by the recognition process.
- the receiving side device 640 uses the feature data restored by the feature restoration unit 642 for both the restoration of the target data by the target restoration unit 643 and the recognition process by the recognition unit 644.
- the restoration process of the target data and the expression content of the restored target data are compared with the case where the recognition process is performed using the restored target data after the target data is restored.
- the processing time for performing the recognition process is short.
- FIG. 26 is a flowchart showing an example of a processing procedure in the information processing method according to the sixth embodiment.
- the processes shown in FIG. 26 include acquiring communication data (step S611), restoring feature data (step S612), restoring target data (step S613), and performing recognition processing (step S613). S614) and outputting the result (step S615).
- step S611 In acquiring the communication data (step S611), the communication data based on the feature data indicating the feature of the expression content of the target data is received. Restoring the feature data (step S612) restores the feature data based on the received communication data. In restoring the target data (step S613), the target data is restored based on the restored feature data. In performing the recognition process (step S614), the recognition process for the expression content of the target data is performed based on the restored feature data. In outputting the result (step S615), information indicating the expression content of the restored target data and the recognition result by the recognition process is output.
- the feature data restored in step S612 is used for both the restoration of the target data in step S613 and the recognition process in step S614.
- the restoration process of the target data and the representation of the restored target data are performed.
- the processing time for performing the recognition processing for the content can be shortened.
- FIG. 27 is a schematic block diagram showing the configuration of a computer according to at least one embodiment.
- the computer 700 includes a CPU (Central Processing Unit) 710, a main storage device 720, an auxiliary storage device 730, and an interface 740.
- CPU Central Processing Unit
- Any one or more or a part of the side devices 640 may be mounted on the computer 700.
- the operation of each of the above-mentioned processing units is stored in the auxiliary storage device 730 in the form of a program.
- the CPU 710 reads the program from the auxiliary storage device 730, expands it to the main storage device 720, and executes the above processing according to the program.
- the CPU 710 secures a storage area corresponding to each of the above-mentioned storage units in the main storage device 720 according to the program. Communication between each device and other devices is executed by having the interface 740 have a communication function and performing communication according to the control of the CPU 710.
- the operations of the feature extraction unit 12, the communication data generation unit 13, and each unit thereof are stored in the auxiliary storage device 730 in the form of a program.
- the CPU 710 reads the program from the auxiliary storage device 730, expands it to the main storage device 720, and executes the above processing according to the program.
- the CPU 710 secures a storage area for processing of the transmitting side device 10 in the main storage device 720 according to the program.
- the acquisition of image data by the image acquisition unit 11 is executed, for example, by having the interface 740 provided with an image pickup device and performing image pickup under the control of the CPU 710.
- Data transmission by the transmission unit 16 is executed when the interface 740 has a communication function and operates according to the control of the CPU 710.
- the operation of the feature restoration unit 22, the acquired image restoration unit 26, the recognition unit 27, and each unit thereof is stored in the auxiliary storage device 730 in the form of a program.
- the CPU 710 reads the program from the auxiliary storage device 730, expands it to the main storage device 720, and executes the above processing according to the program.
- the CPU 710 secures a storage area for processing of the receiving side device 20 in the main storage device 720 according to the program.
- Data reception by the receiving unit 21 is executed when the interface 740 has a communication function and operates according to the control of the CPU 710.
- the output of information by the output unit 28 is executed, for example, by the interface 740 including a display device and displaying an image under the control of the CPU 710.
- the operations of the feature extraction unit 12, the communication data generation unit 31, and each unit thereof are stored in the auxiliary storage device 730 in the form of a program.
- the CPU 710 reads the program from the auxiliary storage device 730, expands it to the main storage device 720, and executes the above processing according to the program.
- the CPU 710 secures a storage area for processing of the transmitting side device 30 such as the noisy feature data storage unit 35 in the main storage device 720 according to the program.
- the acquisition of image data by the image acquisition unit 11 is executed, for example, by having the interface 740 provided with an image pickup device and performing image pickup under the control of the CPU 710.
- Data transmission by the transmission unit 16 is executed when the interface 740 has a communication function and operates according to the control of the CPU 710.
- the operations of the acquired image restoration unit 26, the recognition unit 27, the feature restoration unit 41, and each unit thereof are stored in the auxiliary storage device 730 in the form of a program.
- the CPU 710 reads the program from the auxiliary storage device 730, expands it to the main storage device 720, and executes the above processing according to the program.
- the CPU 710 secures a storage area for processing of the receiving side device 40 such as the noisy feature data storage unit 43 in the main storage device 720 according to the program.
- Data reception by the receiving unit 21 is executed when the interface 740 has a communication function and operates according to the control of the CPU 710.
- the output of information by the output unit 28 is executed, for example, by the interface 740 including a display device and displaying an image under the control of the CPU 710.
- the operations of the feature restoration unit 612, the target restoration unit 613, and the recognition unit 614 are stored in the auxiliary storage device 730 in the form of a program.
- the CPU 710 reads the program from the auxiliary storage device 730, expands it to the main storage device 720, and executes the above processing according to the program.
- the CPU 710 secures a storage area for processing of the information processing device 610 in the main storage device 720 according to the program.
- Data reception by the receiving unit 611 is executed when the interface 740 has a communication function and operates according to the control of the CPU 710.
- the output of information by the output unit 615 is executed, for example, by the interface 740 including a display device and displaying an image under the control of the CPU 710.
- the operations of the feature extraction unit 632 and the communication data generation unit 633 are stored in the auxiliary storage device 730 in the form of a program.
- the CPU 710 reads the program from the auxiliary storage device 730, expands it to the main storage device 720, and executes the above processing according to the program.
- the CPU 710 secures a storage area for processing of the transmitting side device 630 in the main storage device 720 according to the program.
- the acquisition of the target data by the data acquisition unit 631 is executed when the interface 740 includes a device for acquiring the target data such as an image pickup device and operates according to the control of the CPU 710.
- Data transmission by the transmission unit 634 is executed when the interface 740 has a communication function and operates according to the control of the CPU 710.
- the operations of the feature restoration unit 642, the target restoration unit 643, and the recognition unit 644 are stored in the auxiliary storage device 730 in the form of a program.
- the CPU 710 reads the program from the auxiliary storage device 730, expands it to the main storage device 720, and executes the above processing according to the program.
- the CPU 710 secures a storage area for processing of the receiving side device 640 in the main storage device 720 according to the program.
- Data reception by the receiving unit 641 is executed when the interface 740 has a communication function and operates according to the control of the CPU 710.
- the output of information by the output unit 645 is executed, for example, by the interface 740 including a display device and displaying an image under the control of the CPU 710.
- the transmitting side device 10 The transmitting side device 10, the receiving side device 20, the transmitting side device 30, the receiving side device 40, the transmitting side device 51, the receiving side device 52, the setting update device 53, the information processing device 610, the transmitting side device 630, and the receiving side device 630.
- a program for executing all or a part of the processing performed by the side device 640 is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read by the computer system and executed. Processing may be performed.
- the term "computer system” as used herein includes hardware such as an OS (Operating System) and peripheral devices.
- the "computer-readable recording medium” is a portable medium such as a flexible disk, a magneto-optical disk, a ROM (Read Only Memory), a CD-ROM (Compact Disc Read Only Memory), or a hard disk built in a computer system. It refers to a storage device such as.
- the above-mentioned program may be a program for realizing a part of the above-mentioned functions, and may be a program for realizing the above-mentioned functions in combination with a program already recorded in the computer system.
- the receiving means receives the communication data based on the quantized feature data and receives the communication data.
- the feature restoration means includes a dequantization means that dequantizes the quantized feature data based on sampling according to the probability distribution of the feature data before it is quantized.
- the information processing device according to Appendix 1.
- the receiving means expresses the first feature data indicating the characteristics of the expression content of the first target data in the first time step and the second target data in the second time step which is a time step later than the first time step.
- the feature restoration means restores the feature difference data based on the received communication data, and restores the second feature data based on the restored feature difference data and the first feature data.
- the receiving means receives the communication data based on the quantized feature difference data, and receives the communication data.
- the feature restoration means includes a dequantization means that dequantizes the quantized feature difference data based on sampling according to the probability distribution of the feature difference data before being quantized.
- the receiving means receives the communication data based on the first intermediate feature data and the second intermediate feature data calculated based on the data downsampled from the first intermediate feature data.
- the feature restoration means restores the first intermediate feature data based on the data upsampled from the second intermediate feature data restored based on the received communication data.
- the feature restoration means restores the first intermediate feature data by using a process corresponding to the inverse operation of the process of calculating the second intermediate feature data based on the data downsampled from the first intermediate feature data.
- the transmitting side device is Data acquisition means to acquire the target data and A feature extraction means for calculating feature data indicating the features of the expression content of the target data, and A communication data generation means that generates communication data based on the feature data, A transmission means for transmitting the communication data and With
- the receiving device is A receiving means for receiving the communication data and A feature restoration means for restoring the feature data based on the received communication data, and A target restoration means for restoring the target data based on the restored feature data, and a target restoration means.
- a recognition means that performs recognition processing for the expression content of the target data based on the restored feature data, and An output means for outputting information indicating the restored expression content of the target data and the recognition result by the recognition process, and Information processing system equipped with.
- the communication data generation means includes a quantization means for quantizing the feature data.
- the feature restoration means includes a dequantization means that dequantizes the quantized feature data based on sampling according to the probability distribution of the feature data before it is quantized.
- the data acquisition means acquires the first target data in the first time step and the second target data in the second time step, which is a time step later than the first time step.
- the feature extraction means calculates the first feature data indicating the characteristics of the expression content of the first target data and the second feature data indicating the characteristics of the expression content of the second target data.
- the communication data generation means calculates the feature difference data indicating the difference between the first feature data and the second feature data, and generates the communication data based on the calculated feature difference data.
- the feature restoration means restores the feature difference data based on the received communication data, and restores the second feature data based on the restored feature difference data and the first feature data.
- the communication data generation means includes a quantization means for quantizing the feature difference data.
- the feature restoration means includes a dequantization means that dequantizes the quantized feature difference data based on sampling according to the probability distribution of the feature difference data before being quantized.
- the information processing system according to Appendix 10.
- the transmitting side device is Further provided with a noisy feature data storage means for storing the noisy feature data which is the feature data including the quantization error.
- the communication data generation means is The first noisy feature data, which is the first feature data including the quantization error, is read out from the noisy feature data storage means, and the feature difference data indicating the difference between the first noisy feature data and the second feature data is calculated.
- Feature difference calculation means and The quantization error is calculated based on the dequantized data after the feature difference data indicating the difference between the first noisy feature data and the second feature data is quantized and the first noisy feature data.
- a feature restoration means for calculating the second noisy feature data which is the second feature data including the noisy feature data and updating the noisy feature data stored in the noisy feature data storage means to the second noisy feature data.
- the feature extraction means calculates the feature data including the first intermediate feature data and the second intermediate feature data calculated based on the data downsampled from the first intermediate feature data.
- the feature restoration means restores the first intermediate feature data based on the data upsampled from the second intermediate feature data restored based on the received communication data.
- the feature restoration means uses a process corresponding to an inverse operation of a process in which the feature extraction means calculates the second intermediate feature data based on the data downsampled from the first intermediate feature data. Restore intermediate feature data, The information processing system according to Appendix 13.
- (Appendix 16) Receiving communication data based on feature data that shows the features of the representation content of the target data, Restoring the feature data based on the received communication data, Restoring the target data based on the restored feature data, Performing recognition processing for the expression content of the target data based on the restored feature data, and Outputting information indicating the restored expression content of the target data and the recognition result by the recognition process, and Information processing methods including.
- the transmitting side device calculates the feature data indicating the feature of the expression content of the target data, and The transmitting side device generates communication data based on the feature data, and When the transmitting device transmits the communication data, When the receiving device receives the communication data, The receiving device restores the feature data based on the received communication data. The receiving device restores the target data based on the restored feature data. The receiving side device performs recognition processing for the expression content of the target data based on the restored feature data, and The receiving device outputs information indicating the restored expression content of the target data and the recognition result by the recognition process. Information processing methods including.
- the present invention may be applied to an information processing device, an information processing system, an information processing method and a recording medium.
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biodiversity & Conservation Biology (AREA)
- Biomedical Technology (AREA)
- Signal Processing (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Image Processing (AREA)
Abstract
Description
動画像データの圧縮では、深層学習ベースの動画像圧縮技術を用いることができる(非特許文献1から非特許文献3参照)。また、画像認識では、物体検出手法を用いて、画像中でターゲット(監視対象)を検出して追跡することが考えられる(非特許文献4参照)。ターゲットの検出結果は、例えば復元された画像中に表示して監視者に提示することができる。
以下では、情報処理システムが、画像データの送受信および画像認識を行う場合を例に説明する。ただし、以下の実施形態における送受信および認識処理の対象は画像データに限定されず、階層的に圧縮および伸長(復元)可能ないろいろなデータとすることができる。例えば、情報処理システムが、音声データの送受信および音声認識を行うようにしてもよい。あるいは、情報処理システムが、LiDAR(Light Detection And Ranging)などの各種計測装置が出力する点群データを送受信および認識処理の対象としていてもよい。
図1は、第一実施形態に係る情報処理システムの構成例を示す概略ブロック図である。図1に示す構成において、情報処理システム1は、送信側装置10と、受信側装置20とを備える。送信側装置10は、画像取得部11と、特徴抽出部12と、通信データ生成部13と、送信部16とを備える。通信データ生成部13は、量子化部14と、符号化部15とを備える。受信側装置20は、受信部21と、特徴復元部22と、取得画像復元部26と、認識部27と、出力部28とを備える。特徴復元部22は、復号部23と、脱量子化部24と、中間特徴生成部25とを備える。
送信側装置10は、画像を取得し、取得した画像をビットストリーム(Bit Stream)等の送信用データに変換して受信側装置20へ送信する。受信側装置20は、送信側装置10から受信したデータから画像を復元し、また、受信画像に対する画像認識を行う。
ただし、情報処理システム1の用途は特定の用途に限定されない。
そこで、受信側装置20は、受信データからの画像復元の過程で生成する中間特徴データを用いて画像認識を行う。これにより、受信データから画像を復元した後、復元画像を用いて画像認識を行う場合よりも、短時間で効率的に処理を行うことができる。
受信側装置20は、情報処理装置の例に該当する。
あるいは、撮像装置が送信側装置10とは別の装置として構成され、画像取得部11が撮像装置から画像データを取得するようにしてもよい。あるいは、画像取得部11が、画像データを記録している記録媒体からが画像データを読み出すようにしてもよい。
画像取得部11は、取得した画像データを特徴抽出部12へ出力する。
画像取得部11は、取得手段の例に該当する。
特徴抽出部12は、特徴抽出手段の例に該当する。
以下では、特徴抽出部12が、逆演算可能な畳み込みニューラルネットワークによる深層学習モデルを用いて構成される場合を例に説明する。逆演算可能な畳み込みニューラルネットワークによる深層学習モデルを、インバーティブル深層畳み込みニューラルネットワークモデル(Invertible Deep Convolutional Neural Network Model)とも称する。ここでいう逆演算は、元の演算と入出力が逆になる演算である。すなわち、逆演算では、元の演算における出力値が逆演算への入力値となる場合に、元の演算における入力値と同じ値を出力する。
図2の例において、特徴抽出部12は、3つの処理ステージ部112と、2つのチャネル分割部113とを備える。これらは、2つの処理ステージ部112の間のそれぞれにチャネル分割部113が1つずつ設けられる配置で直列に接続され、さらに、前処理部111に直列に接続されている。3つの処理ステージ部112を区別する場合、データの流れの上流側から下流側へ順に、符号112-1、112-2、112-3を付す。2つのチャネル分割部113を区別する場合、データの流れの上流側から下流側へ順に、符号113-1、113-2を付す。
ただし、特徴抽出部12が備える処理ステージ部112の個数は1つ以上であればよい。特徴抽出部12が備えるチャネル分割部113の個数は、処理ステージ部112の個数よりも1つ少なくてもよい。
あるいは、画像取得部11が出力する画像データをそのままニューラルネットワークに入力して特徴抽出を行える場合、特徴抽出部12が、前処理部111を備えていなくてもよい。すなわち、前処理部111による前処理は必須ではない。
図2の例において、中間特徴データからチャネル分割されているデータも、特徴データの一種に該当する。
図3は、処理ステージ部112の構成例を示す概略ブロック図である。図3に示す構成において、処理ステージ部112は、ダウンサンプリング部121と、処理ブロック部122とを備える。
Nは1以上の整数であればよい。
例えば、ダウンサンプリング部121が、縦2個×横2個の4つの画素ごとに1つの画素に置き換えることによって、画素数が4分の1の画像に縮小するようにしてもよい。その場合、ダウンサンプリング部121が、4つの画素の画素値のうち最大値を選択するようにしてもよい。あるいは、ダウンサンプリング部121が、4つの画素の画素値の平均を算出して、サイズ縮小後の画像の画素値として用いるようにしてもよい。
ここでいう入力チャネル数は、ダウンサンプリング部121への入力データにおけるチャネルの個数である。出力チャネル数は、ダウンサンプリング部121からの出力データにおけるチャネルの個数である。
畳み込み処理部133は、処理後のデータをグループCおよびグループDの2つのグループに振り分ける。畳み込み処理部133は、グループCに振り分けたデータを乗算部134に出力し、グループDに振り分けたデータを加算部135に出力する。
図2の例のように処理ステージ部112とチャネル分割部113とを交互に設ける構成とすることで、処理ステージ部112およびチャネル分割部113による処理に対する逆処理を比較的簡単な計算で行うことができる。
通信データ生成部13は、通信データ生成手段の例に該当する。
量子化部14は、量子化手段の例に該当する。
ただし、情報処理システム1が用いる符号化方式は、エントロピ符号化方式に限定されない。ビットストリームなど通信に適したデータを生成可能ないろいろな符号化方式を、情報処理システム1に適用することができる。
送信部16と受信部21との間の通信方式は、特定のものに限定されない。例えば、送信部16と受信部21とが無線通信を行うようにしてもよいし、有線で通信を行うようにしてもよい。
受信部21は、受信手段の例に該当する。
特徴復元部22は、特徴復元手段の例に該当する。
上記のように、情報処理システム1が用いる符号化方式は、エントロピ符号化方式に限定されない。受信側装置20が行う復号はエントロピ復号に限定されず、送信側装置10によって符号化されたデータを復号するものであればよい。
脱量子化部24が整数を実数に変換する方法は、特定の方法に限定されない。例えば、脱量子化部24が、特徴データの要素としての実数ベクトルの符号化確率を表す確率分布を予め記憶しておき、この確率分布に基づいてサンプリングを行うようにしてもよい。この場合、特徴データの要素としての実数ベクトルの符号化確率を表す確率分布は、量子化される前の特徴データの確率分布の例に該当する。
脱量子化部24が、特徴データの確率分布を脱量子化に反映させることによって、脱量子化を高精度に行えると期待される。
あるいは、脱量子化部24が、整数の値はそのままとし、整数データから実数データへ、データ形式のみを変更するようにしてもよい。
脱量子化部24は、脱量子化手段の例に該当する。
脱量子化部24による脱量子化は、量子化部14による量子化に対する近似的な逆演算と捉えることができる。
図5の例において、2つの逆処理ステージ部211と、2つのチャネル結合部212とが、交互に配置されて直列に接続されている。2つの逆処理ステージ部211を区別する場合、データの流れの上流側から下流側へ順に、符号211-1、211-2を付す。2つのチャネル結合部212を区別する場合、データの流れの上流側から下流側へ順に、符号212-1、212-2を付す。
チャネル結合部212-2の出力を、ノイジー中間特徴データY3’と表記する。ノイジー中間特徴データY3’は、処理ステージ部112-1が出力する中間特徴データY1が、量子化ノイズを含んで復元されたデータである。
図6の例において、N個の逆処理ブロック部221が直列に接続され、さらに、アップサンプリング部222が直列に接続されている。N個の逆処理ブロック部221を区別する場合、データの流れの上流側から下流側へ順に、符号221-1、・・・、221-Nを付す。
チャネル分割部231は、グループA’に振り分けたデータを減算部233へ出力し、グループB’に振り分けたデータを畳み込み処理部232およびチャネル結合部235へ出力する。
畳み込み処理部232は、畳み込み処理部133と同様の処理を行う。具体的には、畳み込み処理部232は、グループB’のデータの入力を受けて、入力されたデータに対して畳み込み処理を行う。畳み込み処理部133が、入力されたデータに対して畳み込み処理および非線形変換などの一連の処理を行う場合、畳み込み処理部232も、畳み込み処理部133と同様の一連の処理を行う。畳み込み処理部232が、畳み込みニューラルネットワークを用いて構成されていてもよい。
畳み込み処理部232は、グループD’に振り分けたデータを減算部233に出力し、グループC’に振り分けたデータを除算部234に出力する。
チャネル結合部235の処理は、チャネル分割部132が行う処理に対する逆の処理にも該当する。
逆アフィンチャネル変換部236は、アフィンチャネル変換部131の演算の逆演算を行う。
取得画像復元部26は、対象復元手段の例に該当する。取得画像復元部26による取得画像の復元は、特徴復元部22が復元した特徴データに基づいて取得画像データを復元する処理に該当する。
取得画像復元部26の逆処理ステージ部211を、中間特徴生成部25の逆処理ステージ部(図5)と区別する場合、取得画像復元部26の逆処理ステージ部211を、逆処理ステージ部211-3と表記する。逆処理ステージ部211-3は、処理ステージ部112-1(図2)の逆モデルに該当する。
後処理部241は、前処理部111の演算の逆演算を行う。
復元画像は、取得画像に類似する。具体的には、復元画像は、取得画像に量子化ノイズが加わった画像である。
したがって、認識部27が行う画像認識は、特徴復元部22が復元した特徴データに基づいて、取得画像データの表現内容である取得画像に対する認識処理を行うことに該当する。認識部27は、認識手段の例に該当する。
図9の例において、3つの中間特徴処理部251それぞれに1つずつ位置推定処理部254および分類処理部255が接続されている。
中間特徴処理部251が認識対象を検出する方法として、公知の方法を用いることができる。
加算部253の各々は、中間特徴処理部251の出力と、アップサンプリング部252の出力とを画素ごとに足し合わせる。
位置推定処理部254が認識対象の復元画像における位置を検出する方法として公知の方法を用いることができる。
分類処理部255が認識対象をクラス分類する方法として公知の方法を用いることができる。
NMS処理部256が処理を行う方法として、Non-Maximum Suppressionとして公知の方法を用いるようにしてもよい。
だたし、出力部28が、復元画像と認識結果とを出力する方法は、特定の方法に限定されない。
出力部28が、復元画像と認識結果とを別々に出力するようにしてもよい。
出力部28は、出力手段の例に該当する。
次に、特徴抽出部12は、取得画像の特徴データを抽出する(ステップS102)。
次に、量子化部14は、特徴データを量子化する(ステップS103)。
そして、送信部16は、符号化部15が出力するビットストリームを受信側装置20へ送信する(ステップS105)。
ステップS105の後、送信側装置10は、図10の処理を終了する。
図11の処理において、受信部21は、ビットストリームを受信する(ステップS201)。
取得画像復元部26は、ノイジー中間特徴データに基づいて復元画像を生成する(ステップS205)。
また、認識部27は、ノイジー中間特徴データに基づいて画像認識を行い、認識結果を算出する(ステップS206)。
そして、出力部28は、復元画像および認識結果を出力する(ステップS207)。
ステップS207の後、受信側装置20は、図11の処理を終了する。
脱量子化部24が、特徴データの確率分布を脱量子化に反映させることによって、脱量子化を高精度に行えると期待される。
このように、受信側装置20が、異なる画像サイズの特徴データを用いて取得画像データを復元することによって、送信側装置10での画像の圧縮率の調整が比較的容易になる。
これにより、特徴復元部22が、中間特徴データを比較的高精度に復元できると期待される。
図12は、第二実施形態に係る情報処理システムの構成例を示す概略ブロック図である。図2に示す構成において、情報処理システム2は、送信側装置30と、受信側装置40とを備える。送信側装置30は、画像取得部11と、特徴抽出部12と、通信データ生成部31と、送信部16と、ノイジー特徴データ記憶部35と、を備える。通信データ生成部31は、量子化部14と、符号化部15と、脱量子化部32と、特徴差分算出部33と、特徴算出部34とを備える。受信側装置20は、受信部21と、特徴復元部41と、取得画像復元部26と、認識部27と、出力部28と、ノイジー特徴データ記憶部43とを備える。特徴復元部41は、復号部23と、脱量子化部24と、中間特徴生成部25と、特徴算出部42とを備える。
図12に示す情報処理システム2の構成を、図1に示す情報処理システム1と比較すると、動画像を効率的に伝送し処理するための機能部が追加されている。それ以外の点では、情報処理システム2は、情報処理システム1と同様である。
取得画像データのうちの1つを第一取得画像データと称し、第一取得画像の次に撮像される取得画像のデータを第二取得画像データと称する。第一取得画像データは、第一対象データの例に該当する。第二取得画像データは、第二対象データの例に該当する。
一方、通信データ生成部31は、画像取得部11が取得する2つ目以降の画像については、特徴差分データを算出し、算出した特徴差分データに基づいて通信データを生成する。特徴差分データは、特徴抽出部12が算出する2つの特徴データの相違を示すデータである。例えば、通信データ生成部31は、第一特徴データと第二特徴データとの相違を示す特徴差分データを算出し、算出した特徴差分データに基づいて通信データを生成する。
特に、通信データ生成部31は、量子化部14における量子化、および、脱量子化部32における脱量子化によって、量子化ノイズを含むノイジー特徴差分データを生成し、ノイジー特徴差分データに基づいて通信データを生成する。
ノイジー特徴データ記憶部35は、ノイジー特徴データを一時的に記憶する。ノイジー特徴データ記憶部35が記憶するノイジー特徴データは、次の処理におけるノイジー特徴差分データの生成に用いられる。ここでいう次の処理は、画像取得部11が取得する動画像のフレームごとの処理など、画像取得部11が取得する画像ごとの処理のうち、次の画像に対する処理である。
特徴差分算出部33は、ノイジー特徴差分データを算出する。ノイジー特徴差分データは、連続する処理でそれぞれ生成される特徴データと、1つ前の処理で生成されたノイジー特徴データとの差分データである。
受信側装置40は、情報処理装置の例に該当する。
特徴算出部42は、脱量子化部24が復元する、今回の処理におけるノイジー特徴差分データと、ノイジー特徴データ記憶部43が記憶している前回の処理におけるノイジー特徴データとを足し合わせる。これにより、特徴算出部42は、今回の処理におけるノイジー特徴データを算出する。特徴算出部42は、算出したノイジー特徴データを、中間特徴生成部25へ出力する。また、特徴算出部42は、ノイジー特徴データ記憶部43が記憶している前回の処理におけるノイジー特徴データを、特徴算出部42自らが算出した今回の処理におけるノイジー特徴データに更新する。
図13は、特徴差分算出部33がインバーティブル深層畳み込みニューラルネットワークモデルを用いて構成される場合の例を示している。ただし、特徴差分算出部33の構成は、特定のものに限定されない。
差分処理ステージ部311の各々は、時刻ステップtにおける特徴データと時刻ステップt-1におけるノイジー特徴データとの差分を算出する。
図14の例において、N個の差分処理ブロック部321が直列に接続されている。N個の差分処理ブロック部321を区別する場合、データの流れの上流側から下流側へ順に、符号321-1、・・・、321-Nを付す。
チャネル分割部332から畳み込み処理部333へのデータは、グループBに相当するグループのデータである。また、畳み込み処理部333は、ノイジー特徴データ記憶部35が記憶する、時刻ステップt-1におけるノイジー特徴データを取得する。
具体的には、畳み込み処理部333は、結合後のデータに対して畳み込み処理を行う。畳み込み処理部333が、結合後のデータに対して畳み込み処理および非線形変換などの一連の処理を行うようにしてもよい。畳み込み処理部333が、畳み込みニューラルネットワークを用いて構成されていてもよい。
図16は、特徴算出部34がインバーティブル深層畳み込みニューラルネットワークモデルを用いて構成される場合の例を示している。ただし、特徴算出部34の構成は、特定のものに限定されない。
復元処理ステージ部341の各々は、時刻ステップt-1における特徴データと、時刻ステップtにおけるノイジー特徴差分データとに基づいて、時刻ステップtにおけるノイジー特徴データを算出する。
図17の例において、N個の復元処理ブロック部351が直列に接続されている。N個の復元処理ブロック部351を区別する場合、データの流れの上流側から下流側へ順に、符号351-1、・・・、351-Nを付す。
チャネル分割部361から畳み込み処理部362へのデータは、グループBに相当するグループのデータである。また、畳み込み処理部362は、ノイジー特徴データ記憶部35が記憶する、時刻ステップt-1におけるノイジー特徴データを取得する。
具体的には、畳み込み処理部362は、結合後のデータに対して畳み込み処理を行う。畳み込み処理部362が、結合後のデータに対して畳み込み処理および非線形変換などの一連の処理を行うようにしてもよい。畳み込み処理部362が、畳み込みニューラルネットワークを用いて構成されていてもよい。
特徴復元部41は、特徴復元手段の例に該当する。
第一実施形態での量子化された特徴データの脱量子化の場合と同様、脱量子化部24が、量子化される前の特徴差分データの確率分布に従ったサンプリングに基づく脱量子化を行うようにしてもよい。例えば、脱量子化部24が、特徴差分データの要素としての実数ベクトルの符号化確率を表す確率分布を予め記憶しておき、この確率分布に基づいてサンプリングを行うようにしてもよい。
送信側装置30は、ノイジー特徴差分データの算出に、ノイジー特徴データ記憶部35が記憶しているノイジー特徴データを前回のノイジー特徴データ(時刻ステップt-1)として用いる。受信側装置40は、ノイジー特徴差分データからノイジー特徴データ(時刻ステップt)を復元する際に、ノイジー特徴データ記憶部43が記憶している前回のノイジー特徴データ(時刻ステップt-1)を用いる。
受信側装置40が、送信側装置30と同様の、前回のノイジー特徴データを用いて今回のノイジー特徴データを復元することによって、今回のノイジー特徴データを高精度に復元できると期待される。
次に、特徴抽出部12は、取得画像の特徴データを抽出する(ステップS302)。
t=1であると判定した場合(ステップS303:YES)、量子化部14は、特徴データを量子化する(ステップS311)。
次に、送信側装置30は、時刻ステップtがt=1か否かを判定する(ステップS333)。すなわち、送信側装置30は、ステップS332で送信した画像が、1つめの画像か否かを判定する。
ステップS341の後、送信側装置30は、図19の処理を終了する。
具体的には、特徴差分算出部33は、ノイジー特徴データ記憶部35が記憶しているノイジー特徴データを読み出す。このノイジー特徴データは、送信側装置30による図19の処理の前回の実行で得られたものであるから、時刻ステップt-1のノイジー特徴データである。
ステップS321の後、量子化部14は、特徴差分データを量子化する(ステップS322)。
ステップS322の後、処理がステップS331へ進む。
ステップS351の後、特徴算出部34は、ノイジー特徴データを算出し、ノイジー特徴データ記憶部35に記憶させる(ステップS352)。
具体的には、特徴算出部34は、ノイジー特徴データ記憶部35が記憶しているノイジー特徴データ(時刻ステップt-1)を読み出す。そして、特徴算出部34は、ステップS351で脱量子化部32が算出したノイジー特徴差分データ(時刻ステップt)と、ノイジー特徴データ記憶部35から読み出したノイジー特徴データ(時刻ステップt-1)とに基づいて、ノイジー特徴データ(時刻ステップt)を算出する。特徴算出部34は、算出したノイジー特徴データ(時刻ステップt)をノイジー特徴データ記憶部35に記憶させる。
ステップS352の後、送信側装置30は、図19の処理を終了する。
図20のステップS401およびS402は、ビットストリームが特徴データを表す場合と特徴差分データを表す場合とがある点以外は、図11のステップS201およびS202と同様である。ステップS402において、時刻ステップtがt=1の場合は、量子化された特徴データが得られる。一方、t≧2の場合は、量子化された特徴差分データが得られる。
t=1であると受信側装置40判定した場合(ステップS403:YES)、脱量子化部24は、ノイジー特徴データを算出し、ノイジー特徴データ記憶部43に記憶させる(ステップS411)。
具体的には、脱量子化部24は、図11のステップS203の場合と同様、ビットストリームの復号によって得られたデータを脱量子化することによって、ノイジー特徴データを算出する。そして、脱量子化部24は、算出したノイジー特徴データをノイジー特徴データ記憶部43に記憶させる。
ステップS411の後、処理がステップS431へ進む。
ステップS431からS434は、図11のステップS204~S207と同様である。
ステップS434の後、受信側装置40は、図20の処理を終了する。
ステップS422の後、処理がステップS431へ進む。
受信側装置40によれば、特徴差分データに基づく通信データを受信することによって、特徴データに基づく通信データを受信する場合よりも、通信量が少なくて済むと期待される。
脱量子化部24が、特徴差分データの確率分布を脱量子化に反映させることによって、脱量子化を高精度に行えると期待される。
情報処理システム1または情報処理システム2において、通信データの圧縮率を動的に変化させるなど、送信側装置が行う処理の設定を動的に更新するようにしてもよい。その際、受信側装置が行う処理の設定も動的に更新するようにしてもよい。第三実施形態では、その点について説明する。
設定更新部54は、設定更新手段の例に該当する。
これにより、通信データの圧縮率を動的に変化させるなど処理の設定を動的に変化させることができ、かつ、受信側装置52が、特徴データの復元を高精度に行えると期待される。
図22は、第三実施形態に係る情報処理システムの構成の第二例を示す概略ブロック図である。図22に示す構成において、情報処理システム3bでは、設定更新部54が、送信側装置51に設けられている。それ以外の点では、情報処理システム3bは、情報処理システム3aの場合と同様である。
図24は、第四実施形態に係る情報処理装置の構成例を示す概略ブロック図である。図24に示す構成において、情報処理装置610は、受信部611と、特徴復元部612と、対象復元部613と、認識部614と、出力部615と、を備える。
受信部611は、受信手段の例に該当する、特徴復元部612は、特徴復元手段の例に該当する。対象復元部613は、対象復元手段の例に該当する。認識部614は、認識手段の例に該当する。出力部615は、出力手段の例に該当する。
図25は、第五実施形態に係る情報処理システムの構成例を示す概略ブロック図である。図25に示す構成において、情報処理システム620は、送信側装置630と受信側装置640とを備える。送信側装置630はデータ取得部631と、特徴抽出部632と、通信データ生成部633と、送信部634と、を備える。受信側装置640は、受信部641と、特徴復元部642と、対象復元部643と、認識部644と、出力部645と、を備える。
図26は、第六実施形態に係る情報処理方法における処理の手順の例を示すフローチャートである。図26に示す処理は、通信データを取得すること(ステップS611)と、特徴データを復元すること(ステップS612)と、対象データを復元すること(ステップS613)と、認識処理を行うこと(ステップS614)と、結果を出力すること(ステップS615)とを含む。
図27に示す構成において、コンピュータ700は、CPU(Central Processing Unit、中央処理装置)710と、主記憶装置720と、補助記憶装置730と、インタフェース740とを備える。
画像取得部11による画像データの取得は、例えば、インタフェース740が撮像装置を備え、CPU710の制御に従って撮像を行うことで実行される。送信部16によるデータの送信は、インタフェース740が通信機能を有し、CPU710の制御に従って動作することで実行される。
受信部21によるデータの受信は、インタフェース740が通信機能を有し、CPU710の制御に従って動作することで実行される。出力部28による情報の出力は、例えば、インタフェース740が表示装置を備え、CPU710の制御に従って画像を表示することで実行される。
画像取得部11による画像データの取得は、例えば、インタフェース740が撮像装置を備え、CPU710の制御に従って撮像を行うことで実行される。送信部16によるデータの送信は、インタフェース740が通信機能を有し、CPU710の制御に従って動作することで実行される。
受信部21によるデータの受信は、インタフェース740が通信機能を有し、CPU710の制御に従って動作することで実行される。出力部28による情報の出力は、例えば、インタフェース740が表示装置を備え、CPU710の制御に従って画像を表示することで実行される。
受信部611によるデータの受信は、インタフェース740が通信機能を有し、CPU710の制御に従って動作することで実行される。出力部615による情報の出力は、例えば、インタフェース740が表示装置を備え、CPU710の制御に従って画像を表示することで実行される。
データ取得部631による対象データの取得は、インタフェース740が撮像装置など対象データ取得のためのデバイスを備え、CPU710の制御に従って動作することで実行される。送信部634によるデータの送信は、インタフェース740が通信機能を有し、CPU710の制御に従って動作することで実行される。
受信部641によるデータの受信は、インタフェース740が通信機能を有し、CPU710の制御に従って動作することで実行される。出力部645による情報の出力は、例えば、インタフェース740が表示装置を備え、CPU710の制御に従って画像を表示することで実行される。
また、「コンピュータ読み取り可能な記録媒体」とは、フレキシブルディスク、光磁気ディスク、ROM(Read Only Memory)、CD-ROM(Compact Disc Read Only Memory)等の可搬媒体、コンピュータシステムに内蔵されるハードディスク等の記憶装置のことをいう。また上記プログラムは、前述した機能の一部を実現するためのものであってもよく、さらに前述した機能をコンピュータシステムにすでに記録されているプログラムとの組み合わせで実現できるものであってもよい。
また、上記の実施形態の一部または全部は、以下の付記のようにも記載され得るが、以下には限定されない。
対象データの表現内容の特徴を示す特徴データに基づく通信データを受信する受信手段と、
受信された前記通信データに基づいて前記特徴データを復元する特徴復元手段と、
復元された前記特徴データに基づいて前記対象データを復元する対象復元手段と、
復元された前記特徴データに基づいて前記対象データの表現内容に対する認識処理を行う認識手段と、
復元された前記対象データの表現内容と前記認識処理による認識結果とを示す情報を出力する出力手段と、
を備える情報処理装置。
前記受信手段は、量子化された前記特徴データに基づく前記通信データを受信し、
前記特徴復元手段は、量子化された前記特徴データに対して、量子化される前の前記特徴データの確率分布に従ったサンプリングに基づく脱量子化を行う脱量子化手段を備える、
付記1に記載の情報処理装置。
前記受信手段は、第一時刻ステップにおける第一対象データの表現内容の特徴を示す第一特徴データと、前記第一時刻ステップよりも遅い時刻ステップである第二時刻ステップにおける第二対象データの表現内容の特徴を示す第二特徴データとの相違を示す特徴差分データに基づく前記通信データを受信し、
前記特徴復元手段は、受信された前記通信データに基づいて前記特徴差分データを復元し、復元された前記特徴差分データと、前記第一特徴データとに基づいて前記第二特徴データを復元する、
付記1に記載の情報処理装置。
前記受信手段は、量子化された前記特徴差分データに基づく前記通信データを受信し、
前記特徴復元手段は、量子化された前記特徴差分データに対して、量子化される前の前記特徴差分データの確率分布に従ったサンプリングに基づく脱量子化を行う脱量子化手段を備える、
付記3に記載の情報処理装置。
前記受信手段は、第一中間特徴データと、前記第一中間特徴データからダウンサンプリングされたデータに基づいて算出される第二中間特徴データとに基づく前記通信データを受信し、
前記特徴復元手段は、受信された前記通信データに基づいて復元された前記第二中間特徴データからアップサンプリングされたデータに基づいて前記第一中間特徴データを復元する、
付記1から4の何れか一つに記載の情報処理装置。
前記特徴復元手段は、前記第一中間特徴データからダウンサンプリングされたデータに基づいて前記第二中間特徴データを算出する処理の逆演算に該当する処理を用いて、前記第一中間特徴データを復元する、
付記5に記載の情報処理装置。
前記特徴復元手段の処理と前記対象復元手段の処理との組み合わせが、前記通信データの送信元の装置における対象データからの特徴抽出処理の逆演算に該当する処理となるように、前記通信データの送信元の装置が行う処理の設定、前記特徴復元手段が行う処理の設定、または、前記対象復元手段が行う処理の設定の少なくとも何れかを動的に更新する設定更新手段をさらに備える、
付記6に記載の情報処理装置。
送信側装置と受信側装置とを備え、
前記送信側装置は、
対象データを取得するデータ取得手段と、
前記対象データの表現内容の特徴を示す特徴データを算出する特徴抽出手段と、
前記特徴データに基づいて通信データを生成する通信データ生成手段と、
前記通信データを送信する送信手段と、
を備え、
前記受信側装置は、
前記通信データを受信する受信手段と、
受信された前記通信データに基づいて前記特徴データを復元する特徴復元手段と、
復元された前記特徴データに基づいて前記対象データを復元する対象復元手段と、
復元された前記特徴データに基づいて前記対象データの表現内容に対する認識処理を行う認識手段と、
復元された前記対象データの表現内容と前記認識処理による認識結果とを示す情報を出力する出力手段と、
を備える情報処理システム。
前記通信データ生成手段は、前記特徴データを量子化する量子化手段を備え、
前記特徴復元手段は、量子化された前記特徴データに対して、量子化される前の前記特徴データの確率分布に従ったサンプリングに基づく脱量子化を行う脱量子化手段を備える、
付記8に記載の情報処理システム。
前記データ取得手段は、第一時刻ステップにおける第一対象データと、前記第一時刻ステップよりも遅い時刻ステップである第二時刻ステップにおける第二対象データとを取得し、
前記特徴抽出手段は、前記第一対象データの表現内容の特徴を示す第一特徴データと、前記第二対象データの表現内容の特徴を示す第二特徴データとを算出し、
前記通信データ生成手段は、前記第一特徴データと前記第二特徴データとの相違を示す特徴差分データを算出し、算出した特徴差分データに基づいて前記通信データを生成し、
前記特徴復元手段は、受信された前記通信データに基づいて前記特徴差分データを復元し、復元された前記特徴差分データと、前記第一特徴データとに基づいて前記第二特徴データを復元する、
付記8に記載の情報処理システム。
前記通信データ生成手段は、前記特徴差分データを量子化する量子化手段を備え、
前記特徴復元手段は、量子化された前記特徴差分データに対して、量子化される前の前記特徴差分データの確率分布に従ったサンプリングに基づく脱量子化を行う脱量子化手段を備える、
付記10に記載の情報処理システム。
前記送信側装置は、
量子化誤差を含む前記特徴データであるノイジー特徴データを記憶するノイジー特徴データ記憶手段
をさらに備え、
前記通信データ生成手段は、
量子化誤差を含む前記第一特徴データである第一ノイジー特徴データを前記ノイジー特徴データ記憶手段から読み出し、前記第一ノイジー特徴データと前記第二特徴データとの相違を示す前記特徴差分データを算出する特徴差分算出手段と、
前記第一ノイジー特徴データと前記第二特徴データとの相違を示す前記特徴差分データが量子化された後脱量子化されたデータと、前記第一ノイジー特徴データとに基づいて、量子化誤差を含む前記第二特徴データである第二ノイジー特徴データを算出し、前記ノイジー特徴データ記憶手段が記憶する前記ノイジー特徴データを前記第二ノイジー特徴データに更新する特徴復元手段と、
を備える、
付記11に記載の情報処理システム。
前記特徴抽出手段は、第一中間特徴データと、前記第一中間特徴データからダウンサンプリングされたデータに基づいて算出される第二中間特徴データとを含む前記特徴データを算出し、
前記特徴復元手段は、受信された前記通信データに基づいて復元された前記第二中間特徴データからアップサンプリングされたデータに基づいて前記第一中間特徴データを復元する、
付記8から12の何れか一つに記載の情報処理システム。
前記特徴復元手段は、前記特徴抽出手段が前記第一中間特徴データからダウンサンプリングされたデータに基づいて前記第二中間特徴データを算出する処理の逆演算に該当する処理を用いて、前記第一中間特徴データを復元する、
付記13に記載の情報処理システム。
前記特徴復元手段の処理と前記対象復元手段の処理との組み合わせが、前記通信データの送信元の装置における対象データからの特徴抽出処理の逆演算に該当する処理となるように、前記通信データの送信元の装置が行う処理の設定、前記特徴復元手段が行う処理の設定、または、前記対象復元手段が行う処理の設定の少なくとも何れかを動的に更新する設定更新手段をさらに備える、
付記14に記載の情報処理システム。
対象データの表現内容の特徴を示す特徴データに基づく通信データを受信することと、
受信された前記通信データに基づいて前記特徴データを復元することと、
復元された前記特徴データに基づいて前記対象データを復元することと、
復元された前記特徴データに基づいて前記対象データの表現内容に対する認識処理を行うことと、
復元された前記対象データの表現内容と前記認識処理による認識結果とを示す情報を出力することと、
を含む情報処理方法。
送信側装置が、対象データを取得することと、
前記送信側装置が、前記対象データの表現内容の特徴を示す特徴データを算出することと、
前記送信側装置が、前記特徴データに基づいて通信データを生成することと、
前記送信側装置が、前記通信データを送信することと、
受信側装置が、前記通信データを受信することと、
前記受信側装置が、受信された前記通信データに基づいて前記特徴データを復元することと、
前記受信側装置が、復元された前記特徴データに基づいて前記対象データを復元することと、
前記受信側装置が、復元された前記特徴データに基づいて前記対象データの表現内容に対する認識処理を行うことと、
前記受信側装置が、復元された前記対象データの表現内容と前記認識処理による認識結果とを示す情報を出力することと、
を含む情報処理方法。
コンピュータに、
対象データの表現内容の特徴を示す特徴データに基づく通信データを受信することと、
受信された前記通信データに基づいて前記特徴データを復元することと、
復元された前記特徴データに基づいて前記対象データを復元することと、
復元された前記特徴データに基づいて前記対象データの表現内容に対する認識処理を行うことと、
復元された前記対象データの表現内容と前記認識処理による認識結果とを示す情報を出力することと、
を実行させるためのプログラムを記録する記録媒体。
10、30、630 送信側装置
11 画像取得部
12、632 特徴抽出部
13、31、633 通信データ生成部
14 量子化部
15 符号化部
16、634 送信部
20、40、640 受信側装置
21、611、641 受信部
22、41、612、642 特徴復元部
23 復号部
24、32 脱量子化部
25 中間特徴生成部
26 取得画像復元部
27、614、644 認識部
28、615、645 出力部
33 特徴差分算出部
34、42 特徴算出部
35、43 ノイジー特徴データ記憶部
111 前処理部
112 処理ステージ部
113、132、231 チャネル分割部
121 ダウンサンプリング部
122 処理ブロック部
131 アフィンチャネル変換部
133、232、362 畳み込み処理部
134 乗算部
135、253 加算部
136、212、235、365 チャネル結合部
211 逆処理ステージ部
221 逆処理ブロック部
222、252、312、342 アップサンプリング部
233、363 減算部
234、364 除算部
236 逆アフィンチャネル変換部
241 後処理部
251 中間特徴処理部
254 位置推定処理部
255 分類処理部
311 差分処理ステージ部
341 復元処理ステージ部
351 復元処理ブロック部
610 情報処理装置
613、643 対象復元部
631 データ取得部
Claims (10)
- 対象データの表現内容の特徴を示す特徴データに基づく通信データを受信する受信手段と、
受信された前記通信データに基づいて前記特徴データを復元する特徴復元手段と、
復元された前記特徴データに基づいて前記対象データを復元する対象復元手段と、
復元された前記特徴データに基づいて前記対象データの表現内容に対する認識処理を行う認識手段と、
復元された前記対象データの表現内容と前記認識処理による認識結果とを示す情報を出力する出力手段と、
を備える情報処理装置。 - 前記受信手段は、量子化された前記特徴データに基づく前記通信データを受信し、
前記特徴復元手段は、量子化された前記特徴データに対して、量子化される前の前記特徴データの確率分布に従ったサンプリングに基づく脱量子化を行う脱量子化手段を備える、
請求項1に記載の情報処理装置。 - 前記受信手段は、第一時刻ステップにおける第一対象データの表現内容の特徴を示す第一特徴データと、前記第一時刻ステップよりも遅い時刻ステップである第二時刻ステップにおける第二対象データの表現内容の特徴を示す第二特徴データとの相違を示す特徴差分データに基づく前記通信データを受信し、
前記特徴復元手段は、受信された前記通信データに基づいて前記特徴差分データを復元し、復元された前記特徴差分データと、前記第一特徴データとに基づいて前記第二特徴データを復元する、
請求項1に記載の情報処理装置。 - 前記受信手段は、量子化された前記特徴差分データに基づく前記通信データを受信し、
前記特徴復元手段は、量子化された前記特徴差分データに対して、量子化される前の前記特徴差分データの確率分布に従ったサンプリングに基づく脱量子化を行う脱量子化手段を備える、
請求項3に記載の情報処理装置。 - 前記受信手段は、第一中間特徴データと、前記第一中間特徴データからダウンサンプリングされたデータに基づいて算出される第二中間特徴データとを含む前記特徴データに基づく前記通信データを受信し、
前記特徴復元手段は、受信された前記通信データに基づいて復元された前記第二中間特徴データからアップサンプリングされたデータに基づいて前記第一中間特徴データを復元する、
請求項1から4の何れか一項に記載の情報処理装置。 - 前記特徴復元手段は、前記第一中間特徴データからダウンサンプリングされたデータに基づいて前記第二中間特徴データを算出する処理の逆演算に該当する処理を用いて、前記第一中間特徴データを復元する、
請求項5に記載の情報処理装置。 - 前記特徴復元手段の処理と前記対象復元手段の処理との組み合わせが、前記通信データの送信元の装置における対象データからの特徴抽出処理の逆演算に該当する処理となるように、前記通信データの送信元の装置が行う処理の設定、前記特徴復元手段が行う処理の設定、または、前記対象復元手段が行う処理の設定の少なくとも何れかを動的に更新する設定更新手段をさらに備える、
請求項6に記載の情報処理装置。 - 送信側装置と受信側装置とを備え、
前記送信側装置は、
対象データを取得するデータ取得手段と、
前記対象データの表現内容の特徴を示す特徴データを算出する特徴抽出手段と、
前記特徴データに基づいて通信データを生成する通信データ生成手段と、
前記通信データを送信する送信手段と、
を備え、
前記受信側装置は、
前記通信データを受信する受信手段と、
受信された前記通信データに基づいて前記特徴データを復元する特徴復元手段と、
復元された前記特徴データに基づいて前記対象データを復元する対象復元手段と、
復元された前記特徴データに基づいて前記対象データの表現内容に対する認識処理を行う認識手段と、
復元された前記対象データの表現内容と前記認識処理による認識結果とを示す情報を出力する出力手段と、
を備える情報処理システム。 - 対象データの表現内容の特徴を示す特徴データに基づく通信データを受信することと、
受信された前記通信データに基づいて前記特徴データを復元することと、
復元された前記特徴データに基づいて前記対象データを復元することと、
復元された前記特徴データに基づいて前記対象データの表現内容に対する認識処理を行うことと、
復元された前記対象データの表現内容と前記認識処理による認識結果とを示す情報を出力することと、
を含む情報処理方法。 - コンピュータに、
対象データの表現内容の特徴を示す特徴データに基づく通信データを受信することと、
受信された前記通信データに基づいて前記特徴データを復元することと、
復元された前記特徴データに基づいて前記対象データを復元することと、
復元された前記特徴データに基づいて前記対象データの表現内容に対する認識処理を行うことと、
復元された前記対象データの表現内容と前記認識処理による認識結果とを示す情報を出力することと、
を実行させるためのプログラムを記録する記録媒体。
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/JP2021/001240 WO2022153480A1 (ja) | 2021-01-15 | 2021-01-15 | 情報処理装置、情報処理システム、情報処理方法および記録媒体 |
US18/271,642 US20240062507A1 (en) | 2021-01-15 | 2021-01-15 | Information processing device, information processing system, information processing method, and recording medium |
JP2022574987A JPWO2022153480A5 (ja) | 2021-01-15 | 情報処理装置、情報処理システム、情報処理方法およびプログラム |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/JP2021/001240 WO2022153480A1 (ja) | 2021-01-15 | 2021-01-15 | 情報処理装置、情報処理システム、情報処理方法および記録媒体 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022153480A1 true WO2022153480A1 (ja) | 2022-07-21 |
Family
ID=82448090
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2021/001240 WO2022153480A1 (ja) | 2021-01-15 | 2021-01-15 | 情報処理装置、情報処理システム、情報処理方法および記録媒体 |
Country Status (2)
Country | Link |
---|---|
US (1) | US20240062507A1 (ja) |
WO (1) | WO2022153480A1 (ja) |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2018155232A1 (ja) * | 2017-02-23 | 2018-08-30 | ソニー株式会社 | 情報処理装置、情報処理方法、並びにプログラム |
WO2018207334A1 (ja) * | 2017-05-12 | 2018-11-15 | 日本電気株式会社 | 画像認識装置、画像認識方法および画像認識プログラム |
WO2020241074A1 (ja) * | 2019-05-30 | 2020-12-03 | パナソニック インテレクチュアル プロパティ コーポレーション オブ アメリカ | 情報処理方法及びプログラム |
-
2021
- 2021-01-15 WO PCT/JP2021/001240 patent/WO2022153480A1/ja active Application Filing
- 2021-01-15 US US18/271,642 patent/US20240062507A1/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2018155232A1 (ja) * | 2017-02-23 | 2018-08-30 | ソニー株式会社 | 情報処理装置、情報処理方法、並びにプログラム |
WO2018207334A1 (ja) * | 2017-05-12 | 2018-11-15 | 日本電気株式会社 | 画像認識装置、画像認識方法および画像認識プログラム |
WO2020241074A1 (ja) * | 2019-05-30 | 2020-12-03 | パナソニック インテレクチュアル プロパティ コーポレーション オブ アメリカ | 情報処理方法及びプログラム |
Non-Patent Citations (2)
Title |
---|
PRYSTAVKA, PYLYP ET AL.: "Automated Object Recognition System based on Convolutional Autoencoder", 10TH INTERNATIONAL CONFERENCE ON ADVANCED COMPUTER INFORMATION TECHNOLOGIES (ACIT, September 2020 (2020-09-01), pages 830 - 833, XP033833558, DOI: 10.1109/ACIT49673.2020.9208945 * |
ZHANG, WENRUI: "An Image Recognition Algorithm Based on Self-Encoding and Convolutional Neural Network Fusion", 2019 INTERNATIONAL CONFERENCE ON ELECTRONIC ENGINEERING AND INFORMATICS (EEI, November 2019 (2019-11-01), pages 402 - 407, XP033715073, DOI: 10.1109/EEI48997.2019.00093 * |
Also Published As
Publication number | Publication date |
---|---|
US20240062507A1 (en) | 2024-02-22 |
JPWO2022153480A1 (ja) | 2022-07-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
KR20200044652A (ko) | 영상의 주관적 품질을 평가하는 방법 및 장치 | |
WO2018150083A1 (en) | A method and technical equipment for video processing | |
EP3821373B1 (en) | Video processing | |
CN109525859B (zh) | 模型训练、图像发送、图像处理方法及相关装置设备 | |
US20180352207A1 (en) | Methods and Systems for Transmitting Data in a Virtual Reality System | |
JP2015536108A (ja) | 映像圧縮方法 | |
WO2019056898A1 (zh) | 一种编码、解码方法及装置 | |
US20210166434A1 (en) | Signal processing apparatus and signal processing method | |
WO2012147523A1 (ja) | 撮像装置及び画像生成方法 | |
Ayzik et al. | Deep image compression using decoder side information | |
CN111696039A (zh) | 图像处理方法及装置、存储介质和电子设备 | |
KR20200044667A (ko) | Ai 부호화 장치 및 그 동작방법, 및 ai 복호화 장치 및 그 동작방법 | |
KR20230107878A (ko) | 적응적 얼굴 재현 및 얼굴 복원에 기반한 비디오 화상회의 | |
JP4561649B2 (ja) | 画像圧縮装置、画像圧縮プログラム及び画像圧縮方法、hdr画像生成装置、hdr画像生成プログラム及びhdr画像生成方法、並びに画像処理システム、画像処理プログラム及び画像処理方法 | |
US20240037802A1 (en) | Configurable positions for auxiliary information input into a picture data processing neural network | |
US20240161488A1 (en) | Independent positioning of auxiliary information in neural network based picture processing | |
EP1124384A1 (en) | Information processing apparatus for compressed image data using class-classification adaptive processing | |
WO2022153480A1 (ja) | 情報処理装置、情報処理システム、情報処理方法および記録媒体 | |
CN109474825B (zh) | 一种脉冲序列压缩方法及系统 | |
CN109379590B (zh) | 一种脉冲序列压缩方法及系统 | |
CN116847087A (zh) | 视频处理方法、装置、存储介质及电子设备 | |
CN113747242B (zh) | 图像处理方法、装置、电子设备及存储介质 | |
TW202324308A (zh) | 圖像編解碼方法和裝置 | |
KR20220065652A (ko) | Dvs 데이터 및 fbis 이미지의 퓨전 방법, 이를 구현하는 시스템 및 비-일시적 프로그램 저장 장치 | |
CN113409209A (zh) | 图像去模糊方法、装置、电子设备与存储介质 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 21919373 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 2022574987 Country of ref document: JP Kind code of ref document: A |
|
WWE | Wipo information: entry into national phase |
Ref document number: 18271642 Country of ref document: US |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21919373 Country of ref document: EP Kind code of ref document: A1 |