Image processing method
By processing and fusing image features with different resolutions, the problem of incomplete and unclear edge recognition of image cutting algorithms in the prior art is solved, and more accurate target object area recognition and clearer edge definition are achieved.
Patent Information
- Application Number
- CN202410944780.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-15
- Publication Date
- 2024-11-12
AI Technical Summary
When the existing cutout algorithm recognizes the area occupied by the target object, there are problems such as incomplete area and low edge accuracy.
By obtaining image features with different resolutions, processing and fusing image information, using high-resolution features to retain detailed information, low-resolution features to characterize pixel correlation, identifying target objects and background areas, and thus generating more complete and more accurate edges Target image.
It realizes more accurate target object area recognition, avoids information conflicts, improves edge accuracy, and ensures a clear boundary between the target object and the background.
Smart Images

Figure CN118941813A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of image processing, and in particular to an image processing method. Background Art
[0002] Extracting a specific target object from an image (commonly known as cropping) is a common application scenario of image processing technology. For example, in some scenarios, it may be necessary to crop out a person's face from a photo of a person.
[0003] In order to extract the target object from the image, it is necessary to identify the area occupied by the target object in the image. When current cutout algorithms identify the area occupied by the target object, most of them have problems such as incomplete identification of the area or low edge accuracy of the area. Summary of the invention
[0004] To this end, this application discloses the following technical solutions:
[0005] The first aspect of the present application provides an image processing method, comprising:
[0006] Obtaining a first image feature and a second image feature corresponding to the image to be processed, wherein the first image feature has a first resolution, the second image feature has a second resolution, and the first resolution is greater than the second resolution;
[0007] Processing the first image feature according to the second image feature to obtain first image information of the image to be processed;
[0008] fusing the first image information with the second image information of the image to be processed to obtain target image information of the image to be processed, wherein the second image information is obtained by processing the second image feature;
[0009] The second image information is used to indicate a first area, a second area and a third area in the image to be processed, the first area represents an area occupied by a target object, the second area represents an area occupied by a background of the image to be processed, and the third area represents a transition area between the first area and the second area;
[0010] In the fusion, the first image information is used to indicate pixels belonging to the target object and pixels belonging to the background in the third area to obtain the target image information.
[0011] Optionally, the processing the first image feature according to the second image feature to obtain first image information of the image to be processed includes:
[0012] Processing the second image feature having the second resolution to obtain a weight coefficient feature having the first resolution;
[0013] Fusing the first image feature and the weight coefficient feature to obtain a fused image feature;
[0014] The first image information of the image to be processed is determined according to the fused image feature.
[0015] Optionally, the processing the second image feature to obtain a weight coefficient feature having the first resolution includes:
[0016] determining a third image feature according to the second image feature, wherein a resolution of the third image feature is greater than the second resolution;
[0017] Performing pooling processing on the third image feature to obtain a fourth image feature;
[0018] The fourth image feature is subjected to convolution processing and up-sampling processing to obtain a weight coefficient feature having the first resolution.
[0019] Optionally, performing pooling processing on the third image feature to obtain a fourth image feature includes:
[0020] Performing maximum pooling processing and average pooling processing on the third image features in channel dimensions respectively to obtain maximum pooling image features and average pooling image features;
[0021] A fourth image feature is obtained according to the maximum pooling image feature and the average pooling image feature.
[0022] Optionally, processing the second image feature to obtain the second image information includes:
[0023] Performing pooling processing on the second image feature to obtain a pooled image feature;
[0024] The pooled image features and the second image features are fused to obtain the second image information.
[0025] Optionally, performing pooling processing on the second image feature to obtain a pooled image feature includes:
[0026] Processing the second image feature according to a plurality of pooling windows having different sizes to obtain a plurality of primary pooling image features;
[0027] The multiple primary pooled image features and the second image feature are fused to obtain a pooled image feature.
[0028] Optionally, the fusing the pooled image feature and the second image feature to obtain the second image information includes:
[0029] Processing the pooled image features and the second image features through a plurality of convolutional layers connected in series to obtain the second image information;
[0030] Among them, the input data of the first convolutional layer among the multiple convolutional layers is the second image feature, and the output data of the last convolutional layer is used as the second image information; the input data of each convolutional layer except the first convolutional layer includes the output data of the previous convolutional layer and the upsampled pooled image features obtained by upsampling the pooled image features; the resolution of the output data of each convolutional layer is greater than the resolution of the input data.
[0031] Optionally, both the first image information and the second image information are images having the same resolution as that of the image to be processed, and the pixel value of each pixel in the second image information is used to indicate that the pixel belongs to the first area, the second area or the third area; the pixel value in the third area of the first image information is used to indicate that the pixel belongs to the target object or the background;
[0032] The fusing the first image information and the second image information of the image to be processed to obtain target image information of the image to be processed includes:
[0033] The pixels in the second image information that are located in the third area are replaced by pixels in corresponding positions in the first image information to obtain target image information of the image to be processed.
[0034] Optionally, obtaining a first image feature corresponding to the image to be processed includes:
[0035] Performing deep image processing on the image to be processed to sequentially obtain N levels of image features, wherein the N levels of image features have different resolutions;
[0036] The first image feature is obtained by fusing the image to be processed, the x-th level image feature and the x+1-th level image feature, the resolution of the x-th level image feature and the resolution of the x+1-th level image feature are not less than the resolution of the first image feature, N and x are positive integers, and x+1 is less than N.
[0037] Optionally, the fusing the image to be processed, the xth level image feature and the x+1th level image feature to obtain the first image feature includes:
[0038] Performing down-sampling processing on the image to be processed to obtain a down-sampled image with the first resolution;
[0039] Downsampling the x-th level image features to obtain downsampled x-th level image features having the first resolution;
[0040] The downsampled image, the downsampled x-th level image feature and the x+1-th level image feature are fused to obtain a first image feature. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.
[0042] Figure 1 is a flow chart of an image processing method provided by an embodiment of the present application;
[0043] Figure 2 It is a schematic diagram of the architecture of an image processing model provided in an embodiment of the present application;
[0044] Figure 3 It is a schematic diagram of the architecture of a feature fusion module provided in an embodiment of the present application;
[0045] Figure 4 Schematic diagram of the architecture of a pyramid pooling module provided in an embodiment of the present application;
[0046] Figure 5 is a flow chart of a method for obtaining a first image feature provided by an embodiment of the present application;
[0047] Figure 6 It is a schematic diagram of the effect of an image processing method provided by an embodiment of the present application;
[0048] Figure 7 is a structural schematic diagram of an image processing device provided in an embodiment of the present application;
[0049] Figure 8 It is a structural schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0050] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0051] This application embodiment provides an image processing method, see Figure 1 , is a flow chart of the method, and the method may include the following steps.
[0052] S101, obtaining a first image feature and a second image feature corresponding to an image to be processed, wherein the first image feature has a first resolution, the second image feature has a second resolution, and the first resolution is greater than the second resolution.
[0053] The image to be processed can be any image from which the target object needs to be extracted. The target object can be a human body part in the image to be processed, such as a face, torso, limbs, etc., or can be any object in the image to be processed, such as a vehicle, a house, an animal, etc.
[0054] As an example, the information obtained in S101 may be Figure 2 The first image feature and the second image feature of the image to be processed are shown, and the target object can be Figure 2 The face of the person in the image to be processed is shown.
[0055] The image to be processed can be an image of any resolution or an image with a specific resolution. When the resolution of the original image from which the target object needs to be extracted is not a specific resolution, the original image can be reduced, enlarged, cropped, filled, etc. to convert the original image into an image to be processed with the specific resolution.
[0056] As an example, the image to be processed can be an image of 512 pixels * 512 pixels. When the resolution of the original image is greater than 512 pixels * 512 pixels, the original image can be reduced or the edge of the original image can be cropped to obtain an image to be processed that contains the target object and has a resolution of 512 pixels * 512 pixels.
[0057] The image processing method of this embodiment can be implemented based on a pre-built image processing model. The image processing model can have the following features: Figure 2 The architecture shown.
[0058] The image processing model may include an encoding network, which may include multiple convolutional layers connected in series. In S101, the image to be processed may be input into the encoding network, and the image to be processed may be processed in sequence using the multiple convolutional layers connected in series in the encoding network to obtain multi-level image features of the image to be processed, and then the first image feature and the second image feature may be determined based on the obtained multi-level image features.
[0059] The input data of the first convolutional layer in the encoding network can be the image to be processed, and the input data of each convolutional layer except the first convolutional layer can be the output data of the previous convolutional layer. The output data of each convolutional layer can be used as the first-level image features of the image to be processed. The output data of the first convolutional layer is used as the first-level image features, the output data of the second convolutional layer is used as the second-level image features, and so on.
[0060] This embodiment does not limit the network structure of the coding network and the number of convolutional layers. As some examples, the coding network may be a high-resolution representation learning Net (HRNet), such as HRNet-W18, where W18 indicates that the number of channels in each layer of the network is 18, and the coding network may include 5 convolutional layers. Correspondingly, after the image to be processed is processed by the coding network, the following can be obtained: Figure 2 The 1st to 5th level image features of the image to be processed are shown.
[0061] The resolution of the input data of each convolutional layer can be twice the resolution of its output data, that is, the resolution of the first-level image features can be 1 / 2 of the image to be processed, the resolution of the second-level image features can be 1 / 4 of the image to be processed, the resolution of the third-level image features can be 1 / 8 of the image to be processed, the resolution of the fourth-level image features can be 1 / 16 of the image to be processed, and the resolution of the fifth-level image features can be 1 / 32 of the image to be processed; taking the resolution of the image to be processed as 512*512 as an example, the resolutions of the first to fifth-level image features can be 256*256, 128*128, 64*64, 32*32, and 16*16, respectively.
[0062] After obtaining the multi-level image features, the first-level image features with smaller resolution can be directly used as the second image features, and the first-level image features with larger resolution can be used as the first image features. Figure 2 The 5th level image feature is used as the second image feature, and the 2nd level image feature is used as the first image feature.
[0063] S102: Process the first image feature according to the second image feature to obtain first image information of the image to be processed.
[0064] The image processing model of this embodiment may include two decoding networks, wherein the first decoding network is used to process the first image feature to obtain the first image information, and the second decoding network is used to process the second image feature to obtain the second image information.
[0065] In S102, the second image feature and the first image feature can be directly input into the first decoding network together, and the first decoding network processes the first image feature according to the second image feature to obtain the corresponding first image information. Alternatively, the second image feature can be processed first, and the processed data can be input into the first decoding network, and the first decoding network processes the first image feature according to the processed data to obtain the first image information.
[0066] S103, fusing the first image information and the second image information of the image to be processed to obtain target image information of the image to be processed, wherein the second image information is obtained by processing features of the second image.
[0067] The second image information may be obtained by inputting the second image feature into a second decoding network of the image processing model, processing the second image feature through the second decoding network, and obtaining the corresponding second image information.
[0068] In the fusion of step S103, the first image information is used to indicate pixels belonging to the target object and pixels belonging to the background in the third area to obtain target image information.
[0069] Among them, the second image information is used to indicate the first area, the second area and the third area in the image to be processed, the first area represents the area occupied by the target object, the second area represents the area occupied by the background of the image to be processed, and the third area represents the transition area between the first area and the second area.
[0070] That is to say, the second image information can divide the image to be processed into a first area, a second area and a third area, wherein the first area can be regarded as an area belonging to the target object determined according to the second image feature; the second area can be regarded as an area belonging to the background determined according to the second image feature; and the third area can be regarded as a transition area between the first area and the second area, which cannot be accurately determined as belonging to the background or the target object according to the second image feature.
[0071] The first image information is used to indicate pixels belonging to the target object and pixels belonging to the background in the third area. That is, for each pixel in the third area of the image to be processed, the first image information can indicate whether the pixel belongs to the target object or the background, and based on the first image information, a pixel-level dividing line between the target object and the background can be determined in the image to be processed.
[0072] The target image information may indicate whether each pixel in the image to be processed belongs to the target object or the background. The area consisting of all pixels belonging to the target object indicated by the target image information is equivalent to the area occupied by the target object in the image to be processed.
[0073] The pixels belonging to the target object indicated by the target image information may include:
[0074] Each pixel located in the first area indicated by the second image information, and each pixel located in the third area indicated by the second image information and indicated by the first image information as belonging to the target object.
[0075] The pixels belonging to the target object indicated by the target image information may include:
[0076] Each pixel located in the second area indicated by the second image information, and each pixel located in the third area indicated by the second image information and indicated by the first image information as belonging to the background.
[0077] The first image information and the second image information can be represented in a variety of different forms. As an example, see Figure 2 The first image information and the second image information may both be images with the same resolution as the image to be processed, and the pixel value of each pixel in the second image information is used to indicate that the pixel belongs to the first area, the second area or the third area; the pixel value in the third area in the first image information is used to indicate that the pixel belongs to the target object or the background.
[0078] In this embodiment, the first image information and the second image information can both be grayscale images with the same resolution as the image to be processed, wherein the pixel value of a pixel can be a grayscale value. For example, if the image to be processed is an image of 512 pixels * 512 pixels, then the first image information and the second image information are both grayscale images of 512 pixels * 512 pixels.
[0079] In this case, the first image information, the second image information and the pixels of the image to be processed all correspond one to one.
[0080] The grayscale value of the pixel in the second image information can be between 0 and 255. For any pixel of the second image information, if the grayscale value of the pixel is 255 (that is, the pixel is pure white), it indicates that the pixel belongs to the first area. If the grayscale value of the pixel is 0 (that is, the pixel is pure black), it indicates that the pixel belongs to the second area. If the grayscale value of the pixel is greater than 0 and less than 255 (that is, the color of the pixel is between pure black and pure white), it indicates that the pixel belongs to the third area.
[0081] The grayscale value of a pixel in the first image information can be between 0 and 255, wherein, for any pixel in the first image information located within the third area, the grayscale value of the pixel can be 0 or 255. If the grayscale value of the pixel is 255, it indicates that the pixel belongs to the target object. If the grayscale value of the pixel is 0, it indicates that the pixel belongs to the background. For any pixel in the first image information located outside the third area, the grayscale value of the pixel can be any grayscale value between 0 and 255.
[0082] According to the first image information and the second image information, the method of fusing the two in step S103 may be:
[0083] The pixels in the second image information that are located in the third area are replaced by pixels in corresponding positions in the first image information to obtain target image information of the image to be processed.
[0084] In the above fusion method, pixels in the second image information whose grayscale values are greater than 0 and less than 255 can be determined one by one. For each pixel in the second image information that meets this condition, the pixel is replaced by the pixel at the corresponding position in the first image information. For example, assuming that the grayscale value of the pixel at (x0, y0) in the second image information is 100, the pixel at (x0, y0) in the first image information can be used to replace the pixel at this position in the second image information. After the replacement, the grayscale value of the pixel at (x0, y0) in the second image information becomes 0 or 255, indicating that the pixel at (x0, y0) belongs to the background or the target object.
[0085] The grayscale image obtained after replacing each pixel whose grayscale value is greater than 0 and less than 255 in the second image information in the above manner can be used as the target image information in S103. Figure 2 After the first image information and the second image information are fused in the above manner, we can get Figure 2 The target image information is shown in black and white.
[0086] After the replacement is completed, the grayscale value of each pixel in the target image information can be 0 or 255, wherein a pixel with a grayscale value of 0 indicates that the pixel at the corresponding position in the image to be processed belongs to the background, and a pixel with a grayscale value of 255 indicates that the pixel at the corresponding position in the image to be processed belongs to the target object. Therefore, the area composed of all pixels with a grayscale value of 255 in the target image information can be regarded as the area occupied by the target object in the image to be processed. The second image information is obtained by processing the second image feature, and the target image information represents the area occupied by the target object in the image to be processed.
[0087] The beneficial effects of this embodiment are:
[0088] On the first aspect, the first image feature with a larger resolution can retain more detail information in the image to be processed, and based on the detail information, the boundary between different parts of the image to be processed (such as the background and the target object) can be more accurately determined. The second image feature with a smaller resolution can more accurately characterize the correlation between pixels in the image to be processed, and based on the correlation between pixels, it can be more accurately determined which areas of the image to be processed belong to the target object and which areas belong to the background. This embodiment uses the above characteristics to obtain first image information that can more accurately characterize the boundary between the target object and the background in the transition area based on the first image feature, and obtain second image information that can more accurately reflect the area to which the target object and the background belong based on the second image feature. Finally, the first image information and the second image information are fused to obtain a more complete area occupied by the target object with more precise edges.
[0089] On the second aspect, when determining the first image information, this embodiment not only utilizes the first image feature, but also introduces the second image feature, so that when determining the first image information, the information of the background and the area to which the target object belongs reflected by the second image feature can be referred to, thereby facilitating obtaining more accurate first image information and avoiding the situation where the first image information and the second image information conflict with each other. For example, the following situation can be avoided: for a pixel at the same position, the first image information indicates that it belongs to the target object, while the second image information indicates that it belongs to the second area (i.e., the area of the background).
[0090] Optionally, the manner of processing the first image feature according to the second image feature in step S102 may include:
[0091] Processing the second image feature with the second resolution to obtain a weight coefficient feature with the first resolution;
[0092] Fusing the first image feature and the weight coefficient feature to obtain a fused image feature;
[0093] The first image information of the image to be processed is determined according to the fused image features.
[0094] See also Figure 2 In the above processing method, the second image feature can be processed by a feature fusion module (FFM) of the image processing model to obtain a weight coefficient feature with a first resolution.
[0095] The weight coefficient feature is composed of multiple weight coefficients, and each weight coefficient corresponds to the first feature data of the corresponding position in the first image feature. For example, assuming that the first resolution is 1 / 4 of the resolution of the image to be processed, and the resolution of the image to be processed is 512*512, then the first resolution is 128*128, then the weight coefficient feature can include 128*128 weight coefficients, and the first image feature can include 128*128*N first feature data, where N is the number of channels of the first image feature, and N can be any integer. If the 128*128 first feature data of a channel is regarded as a 128*128 feature matrix, then the first image feature can be regarded as a collection of N 128*128 feature matrices.
[0096] The method of fusing the first image feature and the weight coefficient feature can be that, for each first feature data in the first image feature, the first feature data is multiplied by the weight coefficient of the corresponding position in the weight coefficient feature, and the product obtained is used as the fused feature data of the corresponding position in the fused image feature. After processing each first feature data in the first image feature in this way, all the obtained fused feature data can form a fused image feature.
[0097] Combined with the previous example, for the 128*128 weight coefficient feature, the weight coefficient at (1, 1) can be multiplied by the first feature data at the (1, 1) position on each channel of the first image feature, and the product obtained is used as the fused feature data at the (1, 1) position on the channel in the fused image feature. The weight coefficient at (1, 2) can be multiplied by the first feature data at the (1, 2) position on each channel of the first image feature, and the product obtained is used as the fused feature data at the (1, 2) position on the channel in the fused image feature. And so on. By traversing each first feature data in the first image feature, the fused image feature can be obtained.
[0098] After obtaining the fused image features, the fused image features may be further processed using the first decoding network to obtain the first image information.
[0099] This embodiment does not limit the structure of FFM. As an example, see Figure 3 , FFM can include pooling layers, concatenation layers, convolution layers, activation functions and upsampling layers.
[0100] Among them, when there is only one pooling layer, FFM can contain no concatenation layer, and the output of the pooling layer directly enters the convolution layer. When there are two pooling layers, for example, Figure 3 When using average pooling layers and maximum pooling layers, FFM can contain concatenation layers to merge the outputs of multiple pooling layers.
[0101] The method of processing the second image feature using the FFM may be:
[0102] First, a third image feature is determined according to the second image feature, and a resolution of the third image feature is greater than the second resolution.
[0103] The third image feature here can be obtained by upsampling or convolution processing the second image feature by FFM, or by processing the second image feature by the first convolution layer in the second decoding network.
[0104] The resolution of the third image feature may be twice the second resolution, that is, if the second resolution is 1 / 32 of the resolution of the image to be processed, then the resolution of the third image feature may be 1 / 16 of the resolution of the image to be processed.
[0105] Then, the third image feature is pooled to obtain the fourth image feature.
[0106] The pooling process of the third image feature may be a pooling process in the channel dimension. That is, for each position in the third image feature, the third feature data of each channel at the position in the third image may be obtained, and the obtained multiple feature data may be pooled, and the obtained result may be used as the fourth feature data of the corresponding position in the fourth image feature.
[0107] Exemplarily, assuming that the third image feature includes 3 channels, namely channels 1, 2, and 3, each channel includes 32*32 third feature data. In the above pooling processing link, the data of channel 1, channel 2, and channel 3 located at (1, 1) in the third image feature can be obtained. These 3 data are pooled and the result is obtained as the fourth feature data at the position (1, 1) in the fourth image feature.
[0108] When the FFM includes multiple pooling layers, the fourth image feature may be a result of including outputs of the multiple pooling layers.
[0109] Finally, convolution processing and up-sampling processing are performed on the fourth image feature to obtain a weight coefficient feature with the first resolution.
[0110] In this step, the fourth image feature can be convolved using a convolution layer, and then the result of the convolution is calculated according to an activation function, and then the result of the activation function is upsampled to obtain a weight coefficient feature. The activation function can be a sigmoid function or other types of activation functions.
[0111] Among them, the convolution kernel used for convolution processing in the convolution layer can be a 1*1 convolution kernel, and the upsampling ratio can be 4 times, that is, the resolution of the weight coefficient feature after upsampling is 4 times the resolution of the data before upsampling. Assuming that the resolution of the data before upsampling is 1 / 16 of the resolution of the image to be processed, the resolution of the weight coefficient feature after upsampling is 1 / 4 of the resolution of the image to be processed.
[0112] Optional, such as Figure 3 As shown, the pooling process of the third image feature may include two parts: average pooling process and maximum pooling process. The method of obtaining the fourth image feature may include:
[0113] Performing maximum pooling processing and average pooling processing on the third image feature in the channel dimension respectively to obtain a maximum pooling image feature and an average pooling image feature;
[0114] A fourth image feature is obtained according to the maximum pooling image feature and the average pooling image feature.
[0115] The performing of the maximum pooling process of the channel dimension on the third image feature may include:
[0116] For each position in the third image feature, the third feature data of each channel at that position in the third image can be obtained, and the maximum value among the multiple third feature data obtained is determined as the result of the pooling processing, that is, it is determined as the maximum pooling feature data of the corresponding position in the maximum pooling image feature.
[0117] Exemplarily, assuming that the third image feature includes 3 channels, namely channels 1, 2, and 3, each channel includes 32*32 third feature data. When the third image feature is subjected to maximum pooling processing in the channel dimension, the data of channel 1, the data of channel 2, and the data of channel 3 at (1, 1) in the third image feature can be obtained, and the maximum value of these 3 data is determined as the maximum pooling feature data at the position (1, 1) in the maximum pooling image feature.
[0118] The average pooling process of the channel dimension is performed on the third image feature, which may include:
[0119] For each position in the third image feature, the third feature data of each channel at that position in the third image can be obtained, and the average value of the multiple third feature data obtained is determined as the result of the pooling processing, that is, the average pooling feature data of the corresponding position in the average pooling image feature.
[0120] Exemplarily, assuming that the third image feature includes 3 channels, namely channels 1, 2, and 3, each channel includes 32*32 third feature data. When the third image feature is subjected to average pooling processing in the channel dimension, data of channel 1, data of channel 2, and data of channel 3 at (1, 1) in the third image feature can be obtained, and the average value of these 3 data is determined as the average pooling feature data at the position (1, 1) in the average pooling image feature.
[0121] The way to obtain the fourth image feature based on the maximum pooling image feature and the average pooling image feature can be to splice the maximum pooling image feature and the average pooling image feature in the channel dimension to obtain the fourth image feature. The fourth image feature obtained in this way can include two channels, one channel is the maximum pooling image feature, and the other channel is the average pooling image feature.
[0122] Optionally, the method of processing the second image feature to obtain the second image information may include:
[0123] Performing pooling processing on the second image feature to obtain a pooled image feature;
[0124] The pooled image features and the second image features are fused to obtain second image information.
[0125] like Figure 2As shown, the pooling process of the second image feature can be implemented by the pyramid pooling module of the image processing model.
[0126] In the above processing method, the second image feature can be input into the second decoding network and the pyramid pooling module at the same time, and the pooled image feature output by the pyramid pooling module is then input into the second decoding network, so that the second decoding network combines the second image feature and the pooled image feature to determine the second image information.
[0127] The resolution of the pooled image features may be consistent with the resolution of the second image features.
[0128] The beneficial effects of this embodiment are:
[0129] Pooling the second image features is beneficial to extracting the correlations between pixels at different levels in the image to be processed. When obtaining the second image information, the pooled image features obtained after the second image feature pooling processing are introduced. Then, when determining the second image information, the multi-level correlations between pixels can be referred to. Compared with determining the second image information only based on the single-level correlation reflected by the second image features, referring to the multi-level correlations can obviously obtain more accurate second image information, and thus can more accurately determine the area occupied by the target object.
[0130] This embodiment does not limit the structure of the pyramid pooling module. As an example, Figure 2 The pyramid pooling module can have Figure 4 The structure shown.
[0131] Based on the structure, performing pooling processing on the second image feature to obtain a pooled image feature may include:
[0132] Processing the second image feature according to a plurality of pooling windows having different sizes to obtain a plurality of primary pooling image features;
[0133] A plurality of primary pooled image features and the second image features are fused to obtain a pooled image feature.
[0134] Combination Figure 4 In the above processing method, the second image feature can be input into a pooling window of size (1*1), a pooling window of size (3*3) and a pooling window of size (5*5) respectively, and the second image feature is pooled through these three pooling windows respectively. The output result of each pooling window is processed by subsequent convolution layers and upsampling layers in turn to obtain primary pooling image feature 1, primary pooling image feature 2, and primary pooling image feature 3.
[0135] The upsampling ratio of the upsampling layer can be determined according to the size of the previously connected pooling window to ensure that the primary pooling image features after the upsampling process have the second resolution.
[0136] When the second image feature is pooled according to the pooling window of size (1*1), the pooling can be performed from the channel dimension.
[0137] The method of fusing multiple primary pooled image features and second image features can be to splice multiple primary pooled image features and second image features in the channel dimension to obtain a pooled image feature, or it can be to add the feature data of corresponding positions in the primary pooled image features and the second image features, and use the result as the pooled feature data of the corresponding position in the pooled image feature, thereby obtaining a pooled image feature composed of multiple pooled feature data.
[0138] Optionally, the second decoding network for processing the second image feature may include a plurality of convolutional layers connected in series. In this case, fusing the pooled image feature and the second image feature to obtain the second image information may include:
[0139] Processing the pooled image features and the second image features through a plurality of convolutional layers connected in series to obtain second image information;
[0140] Among them, the input data of the first convolution layer among the multiple convolution layers is the second image feature, and the output data of the last convolution layer is used as the second image information; the input data of each convolution layer except the first convolution layer includes the output data of the previous convolution layer and the upsampled pooled image features obtained by upsampling the pooled image features; the resolution of the output data of each convolution layer is greater than the resolution of the input data.
[0141] Taking the second resolution as 1 / 32 of the resolution of the image to be processed and the second decoding network including 5 convolutional layers as an example, the above processing may include:
[0142] The first convolutional layer in the second decoding network performs convolution processing on the second image features with a resolution of 1 / 32, and outputs image features with a resolution of 1 / 16 of the image to be processed.
[0143] The image features with a resolution of 1 / 16 of the image to be processed output by the first convolutional layer in the second decoding network can be used as the third image features input to the FFM in the aforementioned embodiment.
[0144] The upsampling module performs a 2-fold upsampling process on the pooled image features to obtain upsampled pooled image features with a resolution of 1 / 16 of the image to be processed. Then, the second convolutional layer in the second decoding network can perform a convolution process on the upsampled pooled image features with a resolution of 1 / 16 and the image features with a resolution of 1 / 16 output by the previous convolutional layer, thereby outputting image features with a resolution of 1 / 8 of the image to be processed.
[0145] The upsampling module performs 4 times upsampling on the pooled image features to obtain upsampled pooled image features with 1 / 8 resolution of the image to be processed. Then, the third convolutional layer in the second decoding network can perform convolution on the upsampled pooled image features with 1 / 8 resolution and the image features with 1 / 8 resolution output by the previous convolutional layer, thereby outputting image features with 1 / 4 resolution of the image to be processed.
[0146] The upsampling module performs 8-fold upsampling on the pooled image features to obtain upsampled pooled image features with 1 / 4 resolution of the image to be processed. Then, the fourth convolutional layer in the second decoding network can perform convolution on the upsampled pooled image features with 1 / 4 resolution and the image features with 1 / 4 resolution output by the previous convolutional layer, thereby outputting image features with 1 / 2 resolution of the image to be processed.
[0147] The upsampling module upsamples the pooled image features by a factor of 16 to obtain upsampled pooled image features with 1 / 2 resolution of the image to be processed. Then the fifth convolutional layer in the second decoding network can perform convolution on the upsampled pooled image features with 1 / 2 resolution and the image features with 1 / 2 resolution output by the previous convolutional layer, thereby outputting a second image information with the same resolution as the image to be processed.
[0148] See also Figure 5 , the first image feature corresponding to the image to be processed can be obtained in the following way.
[0149] S501, performing deep image processing on the image to be processed to obtain N levels of image features in sequence, and the resolutions of the N levels of image features are different.
[0150] Combination Figure 2 In the example of step S501, the image to be processed can be processed in sequence by multiple convolutional layers in the encoding network of the image processing model to obtain the first to fifth level image features of the image to be processed.
[0151] S502, fusing the image to be processed, the xth level image feature and the x+1th level image feature to obtain a first image feature, wherein the resolution of the xth level image feature and the resolution of the x+1th level image feature are not less than the resolution of the first image feature.
[0152] Wherein, x and N are both positive integers, and x+1 is less than N.
[0153] Optionally, the resolution of the x+1th level image feature may be consistent with the resolution of the first image feature, and the resolution of the xth level image feature may be greater than the resolution of the first image feature. For example, the resolution of the first image feature may be 1 / 4 of the resolution of the image to be processed, and the resolution of the x+1th level image feature may be Figure 2The second-level image features shown have a resolution of 1 / 4 of the image to be processed, and the x-th-level image features can be Figure 2 The first-level image features shown have a resolution of 1 / 2 of the image to be processed.
[0154] The beneficial effect of this embodiment is that by fusing multi-level image features with a resolution not less than the first resolution and the image to be processed to form a first image feature, a first image feature containing more detail information can be obtained, which is conducive to obtaining more accurate first image information based on the first image feature.
[0155] The method of fusing the image to be processed, the xth level image feature and the x+1th level image feature to obtain the first image feature may include:
[0156] S5021: downsample the image to be processed to obtain a downsampled image with a first resolution.
[0157] The down sampling ratio of the image to be processed may be determined according to a proportional relationship between the first resolution and the resolution of the image to be processed.
[0158] Exemplarily, if the first resolution is 1 / 4 of the resolution of the image to be processed, then in S5021, the image to be processed may be downsampled by 4 times to obtain a downsampled image whose resolution is also 1 / 4 of the image to be processed.
[0159] S5022, down-sample the x-th level image features to obtain down-sampled x-th level image features with a first resolution.
[0160] The magnification of the down-sampling process performed on the x-th level image feature may be determined according to a proportional relationship between the first resolution and the resolution of the x-th level image feature.
[0161] Exemplarily, if the first resolution is 1 / 4 of the resolution of the image to be processed, and the resolution of the x-th level image feature is 1 / 2 of the resolution of the image to be processed, then in S5022, the x-th level image feature can be downsampled by 2 times to obtain a downsampled x-th level image feature whose resolution is also 1 / 4 of the image to be processed.
[0162] S5023, fusing the downsampled image, the downsampled x-th level image features and the x+1-th level image features to obtain a first image feature.
[0163] The fusion method may be to splice the downsampled image, the downsampled x-level image features and the x+1-level image features in the channel dimension to form a multi-channel first image feature. Exemplarily, assuming that the downsampled image, the downsampled x-level image features and the x+1-level image features each include 3 channels, and each channel includes a matrix with a resolution of 1 / 4 of the image to be processed, then after splicing in the channel dimension, a first image feature containing 9 channels can be obtained.
[0164] The fusion method can also be to accumulate the data at corresponding positions in the downsampled image, the downsampled x-th level image features and the x+1-th level image features, and use the result as the first feature data at the corresponding position in the first image feature, thereby obtaining a first image feature composed of multiple first feature data.
[0165] When constructing the image processing model used in this embodiment, the model parameters of the image processing model may be adjusted as follows.
[0166] Obtaining a sample image and sample second image information obtained by annotating the sample image, wherein the sample second image information is used to characterize a first area belonging to the target object, a second area belonging to the background, and a third area serving as a transition in the sample image;
[0167] Process the sample image using the image processing model to obtain predicted second image information corresponding to the sample image;
[0168] According to the difference between the predicted second image information and the sample second image information, a second loss is determined, and the second loss can be a cross entropy loss. According to the second loss, the parameters of the encoding network and the second decoding network in the image processing model are adjusted. According to the second loss, the parameters of the pyramid pooling module can also be adjusted according to the second loss until the second loss converges.
[0169] Obtaining a sample image and sample first image information obtained by annotating the sample image, wherein the sample first image information is used to indicate which pixels in the third area of the sample image belong to the target object and which pixels belong to the background;
[0170] Process the sample image using the image processing model to obtain predicted first image information corresponding to the sample image;
[0171] According to the difference between the predicted first image information and the sample first image information, a first loss is determined, and the parameters of the encoding network and the first decoding network in the image processing model are adjusted according to the first loss. The parameters of the FFM can also be adjusted according to the first loss until the first loss converges.
[0172] The first loss may be the sum of an Alpha loss and a gradient loss determined according to a difference between the predicted first image information and the sample first image information.
[0173] Obtaining a sample image and sample target image information corresponding to the sample image, wherein the sample target image information represents an area occupied by a target object in the sample image;
[0174] Using the image processing model to process the sample image according to the above image processing method, to obtain the predicted target image information of the sample image;
[0175] According to the difference between the predicted target image information and the sample target image information, a third loss is determined, and the parameters of the encoding network in the image processing model are adjusted according to the third loss until the third loss converges.
[0176] Among them, the third loss may include Alpha loss, gradient loss and synthesis loss determined according to the difference between the predicted target image information and the sample target image information. For example, the third loss may be equal to the sum of these three losses.
[0177] For the method of adjusting the parameters of the corresponding parts of the image processing model according to the various losses in the above parameter adjustment process, please refer to the relevant technologies in the field of deep learning and will not be described in detail.
[0178] The sample images used in the above parameter adjustment process may include images in the P3M-10K dataset. Optionally, in order to increase the number and diversity of sample images, the images in the dataset may be randomly flipped, random noise may be added, and regularization may be performed, and the processed images may be used as sample images.
[0179] Optionally, when the image processing model is used to process images of a specific resolution, the images in the data set may be randomly cropped to obtain images of the specific resolution as sample images. For example, the images in the data set may be randomly cropped to obtain sample images of 512*512.
[0180] As some examples, some methods for determining target object areas using the image processing method of this embodiment and other image processing methods in the related technical field are provided below for comparison.
[0181] See also Figure 6 , Figure 6In the figure, (1-1), (2-1) and (3-1) are three images to be processed, (1-2) is a schematic diagram of the area occupied by the target object in (1-1) obtained by manual annotation, where the white part is the area occupied by the target object. Similarly, (2-2) and (3-2) are schematic diagrams of the area occupied by the target objects in (2-1) and (3-1) obtained by manual annotation. In this example, (1-2), (2-2) and (3-2) are used as reference values for evaluating the accuracy of other methods. (1-3), (2-3) and (3-3) are obtained by processing (1-1), (2-1) and (3-1) respectively using other image processing methods in the relevant technical field, and are schematic diagrams of the area occupied by the target objects in the corresponding images.
[0182] (1-4), (2-4) and (3-4) are schematic diagrams of the areas occupied by target objects in the corresponding images obtained by processing (1-1), (2-1) and (3-1) respectively using the image processing method of this embodiment.
[0183] By comparing the parts in the black rectangular frames in (1-2) and (1-3), the parts in the white oval frames in (2-2) and (2-3), and the parts in the black rectangular frames in (3-2) and (3-3), it can be seen that the areas occupied by the target objects determined by other image processing methods have obvious defects relative to the reference values. For example, in (1-3) and (3-3), some pixels that actually belong to the target objects are identified as black pixels representing the background, and in (2-3), some pixels that actually belong to the background are identified as white pixels representing the target object.
[0184] By comparing the parts within the black rectangular frame in (1-2) and (1-4), the parts within the white oval frame in (2-2) and (2-4), and the parts within the black rectangular frame in (3-2) and (3-4), it can be seen that the area occupied by the target object determined by the image processing method of this embodiment is basically consistent with the reference value, and the above-mentioned defects do not exist.
[0185] It can be seen from this that, compared with other image processing methods in the related technical field, the image processing method of this embodiment can more accurately determine the area occupied by the target object in the image to be processed.
[0186] This embodiment also provides an image processing device. Figure 7 , the device may include the following units.
[0187] An obtaining unit 701 is used to obtain a first image feature and a second image feature corresponding to the image to be processed, wherein the first image feature has a first resolution, the second image feature has a second resolution, and the first resolution is greater than the second resolution;
[0188] A processing unit 702 is used to process the first image feature according to the second image feature to obtain first image information of the image to be processed;
[0189] A fusion unit 703 is used to fuse the first image information and the second image information of the image to be processed to obtain target image information of the image to be processed, wherein the second image information is obtained by processing the second image feature;
[0190] The second image information is used to indicate a first area, a second area, and a third area in the image to be processed, the first area indicates an area occupied by the target object, the second area indicates an area occupied by the background of the image to be processed, and the third area indicates a transition area between the first area and the second area;
[0191] The fused first image information is used to indicate pixels belonging to the target object and pixels belonging to the background in the third area to obtain target image information.
[0192] Optionally, the processing unit 702 processes the first image feature according to the second image feature to obtain first image information of the image to be processed, including:
[0193] Processing the second image feature with the second resolution to obtain a weight coefficient feature with the first resolution;
[0194] Fusing the first image feature and the weight coefficient feature to obtain a fused image feature;
[0195] The first image information of the image to be processed is determined according to the fused image features.
[0196] Optionally, the processing unit 702 processes the second image feature to obtain a weight coefficient feature with a first resolution, including:
[0197] determining a third image feature according to the second image feature, wherein a resolution of the third image feature is greater than the second resolution;
[0198] Performing pooling processing on the third image feature to obtain a fourth image feature;
[0199] The fourth image feature is subjected to convolution processing and up-sampling processing to obtain a weight coefficient feature with a first resolution.
[0200] Optionally, the processing unit 702 performs pooling processing on the third image feature to obtain a fourth image feature, including:
[0201] Performing maximum pooling processing and average pooling processing on the third image feature in the channel dimension respectively to obtain a maximum pooling image feature and an average pooling image feature;
[0202] A fourth image feature is obtained according to the maximum pooling image feature and the average pooling image feature.
[0203] Optionally, the processing unit 702 processes the second image feature to obtain the second image information, including:
[0204] Performing pooling processing on the second image feature to obtain a pooled image feature;
[0205] The pooled image features and the second image features are fused to obtain second image information.
[0206] Optionally, the processing unit 702 performs pooling processing on the second image feature to obtain a pooled image feature, including:
[0207] Processing the second image feature according to a plurality of pooling windows having different sizes to obtain a plurality of primary pooling image features;
[0208] A plurality of primary pooled image features and the second image features are fused to obtain a pooled image feature.
[0209] Optionally, the processing unit 702 performs fusion processing on the pooled image feature and the second image feature to obtain second image information, including:
[0210] Processing the pooled image features and the second image features through a plurality of convolutional layers connected in series to obtain second image information;
[0211] Among them, the input data of the first convolution layer among the multiple convolution layers is the second image feature, and the output data of the last convolution layer is used as the second image information; the input data of each convolution layer except the first convolution layer includes the output data of the previous convolution layer and the upsampled pooled image features obtained by upsampling the pooled image features; the resolution of the output data of each convolution layer is greater than the resolution of the input data.
[0212] Optionally, the first image information and the second image information are both images having the same resolution as the resolution of the image to be processed, and the pixel value of each pixel in the second image information is used to indicate that the pixel belongs to the first area, the second area or the third area; the pixel value in the third area in the first image information is used to indicate that the pixel belongs to the target object or the background;
[0213] The fusion unit 703 fuses the first image information and the second image information of the image to be processed to obtain target image information of the image to be processed, including:
[0214] The pixels in the second image information that are located in the third area are replaced by pixels in corresponding positions in the first image information to obtain target image information of the image to be processed.
[0215] Optionally, the obtaining unit 701 obtains a first image feature corresponding to the image to be processed, including:
[0216] Performing deep image processing on the image to be processed sequentially obtains N levels of image features, and the resolutions of the N levels of image features are different;
[0217] The first image feature is obtained by fusing the image to be processed, the x-th level image feature and the x+1-th level image feature, the resolution of the x-th level image feature and the resolution of the x+1-th level image feature are not less than the resolution of the first image feature, N and x are both positive integers, and x+1 is less than N.
[0218] Optionally, the obtaining unit 701 fuses the image to be processed, the xth level image feature and the x+1th level image feature to obtain the first image feature, including:
[0219] Performing down-sampling processing on the image to be processed to obtain a down-sampled image with a first resolution;
[0220] Downsampling the x-th level image features to obtain downsampled x-th level image features with a first resolution;
[0221] The downsampled image, the downsampled x-th level image features and the x+1-th level image features are fused to obtain the first image features.
[0222] The working principle of the image processing device provided in this embodiment can be found in the relevant steps of the image processing method provided in the embodiment of the present application, and will not be repeated here.
[0223] This embodiment also provides an electronic device, see Figure 8 , the electronic device includes a memory 801 and a processor 802;
[0224] Wherein, the memory 801 is used to store computer programs;
[0225] The processor 802 is used to execute the computer program. When the computer program is executed, it is used to implement the image processing method provided in any embodiment of the present application.
[0226] This embodiment also provides a computer storage medium for storing a computer program. When the stored computer program is executed, it is used to implement the image processing method provided in any embodiment of the present application.
[0227] This embodiment also provides a computer program product, which includes a computer program carried on a non-transitory computer-readable medium. When the computer program is executed, it is used to implement the image processing method provided in any embodiment of the present application.
[0228] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.
[0229] For the convenience of description, the above system or device is described by dividing it into various modules or units according to its functions. Of course, when implementing the present application, the functions of each unit can be implemented in the same or multiple software and / or hardware.
[0230] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly contributed to the prior art in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.
[0231] Finally, it should be noted that, in this article, relational terms such as first, second, third and fourth are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.
[0232] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.
Claims
1. An image processing method, include: Obtaining a first image feature and a second image feature corresponding to the image to be processed, wherein the first image feature has a first resolution, the second image feature has a second resolution, and the first resolution is greater than the second resolution; Processing the first image feature according to the second image feature to obtain first image information of the image to be processed; fusing the first image information with the second image information of the image to be processed to obtain target image information of the image to be processed, wherein the second image information is obtained by processing the second image feature; The second image information is used to indicate a first area, a second area and a third area in the image to be processed, the first area represents an area occupied by a target object, the second area represents an area occupied by a background of the image to be processed, and the third area represents a transition area between the first area and the second area; In the fusion, the first image information is used to indicate pixels belonging to the target object and pixels belonging to the background in the third area to obtain the target image information.
2. The method according to claim 1, wherein the first image feature is processed according to the second image feature to obtain the first image information of the image to be processed, include: Processing the second image feature having the second resolution to obtain a weight coefficient feature having the first resolution; Fusing the first image feature and the weight coefficient feature to obtain a fused image feature; The first image information of the image to be processed is determined according to the fused image feature.
3. The method according to claim 2, wherein the second image feature is processed to obtain a weight coefficient feature having the first resolution, include: determining a third image feature according to the second image feature, wherein a resolution of the third image feature is greater than the second resolution; Performing pooling processing on the third image feature to obtain a fourth image feature; The fourth image feature is subjected to convolution processing and up-sampling processing to obtain a weight coefficient feature having the first resolution.
4. The method according to claim 3, wherein the third image feature is subjected to pooling processing to obtain a fourth image feature, include: Performing maximum pooling processing and average pooling processing on the third image features in channel dimensions respectively to obtain maximum pooling image features and average pooling image features; A fourth image feature is obtained according to the maximum pooling image feature and the average pooling image feature.
5. The method according to claim 1, processing the second image feature to obtain the second image information, include: Performing pooling processing on the second image feature to obtain a pooled image feature; The pooled image features and the second image features are fused to obtain the second image information.
6. The method according to claim 5, wherein the second image feature is subjected to pooling processing to obtain a pooled image feature, include: Processing the second image feature according to a plurality of pooling windows having different sizes to obtain a plurality of primary pooling image features; The multiple primary pooled image features and the second image feature are fused to obtain a pooled image feature.
7. The method according to claim 5, wherein the pooled image features and the second image features are fused to obtain the second image information. include: Processing the pooled image features and the second image features through a plurality of convolutional layers connected in series to obtain the second image information; Among them, the input data of the first convolutional layer among the multiple convolutional layers is the second image feature, and the output data of the last convolutional layer is used as the second image information; the input data of each convolutional layer except the first convolutional layer includes the output data of the previous convolutional layer and the upsampled pooled image features obtained by upsampling the pooled image features; the resolution of the output data of each convolutional layer is greater than the resolution of the input data.
8. The method according to claim 1, wherein the first image information and the second image information are both images with the same resolution as the image to be processed, and the pixel value of each pixel in the second image information is used to indicate that the pixel belongs to the first area, the second area or the third area; the pixel value in the third area of the first image information is used to indicate that the pixel belongs to the target object or the background; fusing the first image information and the second image information of the image to be processed to obtain target image information of the image to be processed, include: The pixels in the second image information that are located in the third area are replaced by pixels in corresponding positions in the first image information to obtain target image information of the image to be processed.
9. The method according to claim 1, obtaining a first image feature corresponding to the image to be processed, include: Performing deep image processing on the image to be processed to sequentially obtain N levels of image features, wherein the N levels of image features have different resolutions; The first image feature is obtained by fusing the image to be processed, the x-th level image feature and the x+1-th level image feature, the resolution of the x-th level image feature and the resolution of the x+1-th level image feature are not less than the resolution of the first image feature, N and x are positive integers, and x+1 is less than N.
10. The method according to claim 9, wherein the first image feature is obtained by fusing the image to be processed, the xth level image feature and the x+1th level image feature, include: Performing down-sampling processing on the image to be processed to obtain a down-sampled image with the first resolution; Downsampling the x-th level image features to obtain downsampled x-th level image features having the first resolution; The downsampled image, the downsampled x-th level image feature and the x+1-th level image feature are fused to obtain a first image feature.
Citation Information
Patent Citations
Image processing method and device, electronic equipment and computer readable storage medium
CN110428366A
Image processing method and device, electronic equipment and computer readable storage medium
CN111091521A
Interactive image matting using neural networks
CN111754517A
Image processing method and device, electronic equipment and storage medium
CN113205451A
Portrait matting method and device, computer equipment and readable storage medium
CN113870283A
