Image encoding and decoding method and system for multi-resolution output

Abstract

The present invention discloses an image encoding and decoding method with multi-resolution output, the method comprising: S1: extracting multi-scale features of an image and constructing a multi-scale feature library; S2: based on the multi-scale feature library, training an extreme learning machine model and a support vector machine model to predict the target existence probability and the target size; S3: when the predicted target existence probability is greater than the target existence probability threshold, selecting a multi-scale feature matching the predicted target size in the multi-scale feature library, encoding, and obtaining the encoding feature; S4: performing multi-scale expansion convolution on the encoding feature to obtain a decoded image; S5: based on the decoded image, identifying the target, and correcting the encoding and decoding system based on the identification result. The present invention selects multi-scale features of appropriate resolution for encoding according to the target size category, and corrects the encoding and decoding system according to the decoded image identification result, which is conducive to adapting to scene requirements and improving encoding and decoding efficiency and quality.

Description

Technical Field

[0001] The present invention relates to the field of image coding and decoding technology, and in particular to a method and system for image coding and decoding with multi-resolution output. Background Art

[0002] Image coding and decoding technology is mainly used in image storage or transmission. In the image coding stage, image data is compressed and encoded according to certain rules, which can effectively reduce image storage space and achieve efficient image transmission. At the same time, encoding images can ensure data security and integrity while reducing the amount of calculation and improving image processing efficiency. In the image decoding stage, the encoded image data is restored according to the coding rules and image characteristics, and the restored image is ensured to be undistorted.

[0003] At present, the commonly used image coding and decoding technologies are mainly divided into lossless coding and decoding and lossy coding and decoding. Among them, lossless coding and decoding uses predictive coding and decoding and Huffman coding and decoding to reduce the storage space during image transmission without losing any image information. It is mainly suitable for scenes with high requirements for image quality; lossy coding and decoding removes information that has little impact on image quality, and uses discrete cosine transform coding and decoding and wavelet transform coding and decoding to process images. JPEG is a common lossy coding and decoding method, and PNG is a common lossless coding and decoding method. However, the above methods can only perform quantization coding on images after signal processing and conversion. They cannot adapt to different types of image coding and decoding requirements, and cannot improve the image coding and decoding quality by understanding and reasoning about the image content. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes an image encoding and decoding method and system with multi-resolution output. On the basis of obtaining multi-resolution output of the image, scale features of appropriate resolution are selected for encoding according to the image inference results, and the encoding and decoding system is corrected according to the decoded image recognition results, which is conducive to adapting to scene requirements and improving encoding and decoding efficiency and quality.

[0005] The present invention provides a multi-resolution output image encoding and decoding method, comprising the following steps:

[0006] S1: Extract multi-scale features of the image and build a multi-scale feature library;

[0007] S2: Based on the multi-scale feature library, an extreme learning machine model and a support vector machine model are trained to predict the target existence probability and target size;

[0008] S3: setting a target existence probability threshold, when the predicted target existence probability is greater than the target existence probability threshold, selecting a multi-scale feature matching the predicted target size in the scale feature library, encoding it, and obtaining the encoded feature;

[0009] S4: At the receiving end of the encoding and decoding system, after performing multi-scale expansion convolution on the received encoding features, an expansion matrix group is obtained, and the expansion matrix group is decoded to obtain a decoded image;

[0010] S5: Based on the decoded image, identify the target, and modify the encoding and decoding system based on the identification result.

[0011] Optionally, S1 includes:

[0012] S11: Use multiple convolutional networks to extract multi-scale features from images:

[0013] ;

[0014] in, represents the total number of scales, Indicates the scale number, represents the image to be extracted with multi-scale features, Represents the image In the Convolution operations are performed at different scales. Representing images In the Multi-scale features at different scales;

[0015] S12: Apply the attention mechanism on the multi-scale features to generate weighted multi-scale features;

[0016] S13: The highest resolution value among all weighted multi-scale features is taken as the target resolution ,in represents the target height resolution, Indicates the target width resolution;

[0017] For each weighted multi-scale feature, calculate the scaling ratio of its resolution to the target resolution:

[0018] ;

[0019] in, Indicates the scaling factor of the height resolution, Indicates the scaling factor of the width resolution, represents the height resolution of the weighted multi-scale features, Represents the width resolution of weighted multi-scale features;

[0020] S14: For the weighted multi-scale features, the resolution is expanded by periodic arrangement to obtain multi-scale features with expanded resolution and form a multi-scale feature library:

[0021] ;

[0022] in, represents the multi-scale features obtained after expanding the resolution, Representing scale characteristics Index values ​​in height and width direction, represents the weighted multi-scale features, Indicates rounding down; It is a multi-scale feature library.

[0023] Optionally, the S12 includes:

[0024] S121: Based on the channel attention mechanism, channel weighting is performed on the multi-scale features to obtain the channel-weighted multi-scale features:

[0025] ;

[0026] in, represents the channel weighted feature, represents channel-wise dot product, represents the channel weight; represents the activation function, represents the fully connected layer, represents the global average pooling operation;

[0027] S122: Based on the spatial attention mechanism, multi-scale features are spatially weighted to obtain spatially weighted multi-scale features:

[0028] ;

[0029] in, represents the multi-scale features after spatial weighting, represents the spatial weight; represents the convolution operation, Indicates the maximum pooling of multi-scale features. Indicates average pooling of multi-scale features. Indicates channel splicing, Indicates channel concatenation of maximum pooling and average pooling;

[0030] S123: Based on the channel-weighted multi-scale features and the spatial-weighted multi-scale features, a weighted multi-scale feature is generated:

[0031] ;

[0032] Optionally, the S2 includes:

[0033] S21: Input the multi-scale feature library into the trained extreme learning machine model , predict the probability of target existence;

[0034] S22: Inputting the multi-scale feature library into the trained support vector machine model , predict the target size;

[0035] pass classifiers to identify the target size category and calculate all The scores of the classifiers are calculated, and the target size category corresponding to the classifier with the highest score is taken as the predicted target size.

[0036] Optionally, the S21 includes:

[0037] Input the multi-scale feature library into the trained extreme learning machine model , Output decision function ;

[0038] if , the predicted target does not exist, the predicted target exists probability ;

[0039] if , predicting the existence of the target, the predicted probability of the target existence is expressed as:

[0040] ;

[0041] in, represents the predicted probability of the target existence, represents an exponential function, the probability of the target existence satisfies .

[0042] Optionally, the S3 includes:

[0043] S31: Setting the target existence probability threshold , when the predicted target existence probability is greater than the target existence probability threshold, a multi-scale feature matching the predicted target size is selected in the multi-scale feature library, and detail features are extracted from the selected multi-scale feature:

[0044] ;

[0045] in, represents the selected multi-scale features, Represents the detailed features extracted from the selected multi-scale features, represents the detail feature enhancement parameter, Represents the features obtained after smoothing the selected multi-scale features; represents a two-dimensional Gaussian filter kernel, Represents the convolution operation;

[0046] S32: Combine detail features with selected multi-scale features Fusion, get the encoding features:

[0047] ;

[0048] in, represents the encoding feature, Indicates that and Make stitching.

[0049] Optionally, the S4 includes:

[0050] S41: Perform multi-scale dilation convolution on the coding feature to obtain a dilation matrix group:

[0051] ;

[0052] in, represents the elements of the convolution kernel, represents the element value in the encoded feature, Indicates that the index in the encoded feature is Elements of represents the expansion rate, The values ​​are 1 and 2 respectively. is the expansion matrix, To form an expansion matrix group;

[0053] S42: Decoding the expanded matrix group to obtain a decoded image;

[0054] ;

[0055] in, represents the decoded image obtained, Represents the expansion matrix Bilinear interpolation is used for upsampling. Indicates that and Perform feature fusion, It means that convolution is used to refine the fused features, reduce the number of channels and enhance the representation ability.

[0056] Optionally, S5 includes:

[0057] Identify the target and modify the encoding and decoding system based on the identification results:

[0058] If the target does not exist, it means the target exists probability threshold If the setting is unreasonable, the threshold is updated and the process jumps to step S3 for re-encoding. The target existence probability threshold is updated as follows:

[0059] ;

[0060] in, Indicates the target existence probability threshold update interval;

[0061] If the target exists, but the target recognition size category is inconsistent with the target real size category, it means that the detail feature enhancement parameters Improper settings, increase The value of jumps to step S3 for re-encoding;

[0062] If the target recognition accuracy is less than 60%, it means that the expansion rate If the value is inappropriate, jump to step S41 and update the expansion rate , recalculate the expansion matrix group.

[0063] The present invention also provides an image encoding and decoding system with multi-resolution output, comprising:

[0064] Multi-scale feature module: extracts multi-scale features of images and builds a multi-scale feature library;

[0065] Target existence probability and target size prediction module: Based on the multi-scale feature library, an extreme learning machine model and a support vector machine model are trained to predict the target existence probability and target size;

[0066] Encoding module: setting a target existence probability threshold, when the predicted target existence probability is greater than the target existence probability threshold, selecting a multi-scale feature matching the predicted target size in the multi-scale feature library, encoding it, and obtaining the encoded feature;

[0067] Decoding module: The receiving end of the encoding and decoding system receives the encoding features, performs multi-scale dilation convolution and decoding processing on the encoding features, and obtains a decoded image;

[0068] System correction module: Identify the target based on the decoded image and correct the encoding and decoding system based on the identification result.

[0069] Compared with the prior art, the present invention has the following beneficial effects:

[0070] 1. The present invention accurately obtains multi-resolution output of the image by extracting multi-scale features of the image; on this basis, the extreme learning machine model and the support vector machine model are applied to improve the target existence probability and the target size prediction accuracy; then, the appropriate multi-scale features are selected for encoding according to the target size type, which is conducive to improving the encoding efficiency and encoding quality based on the understanding and reasoning of the image content, and can adapt to different types of image encoding requirements; in the decoding stage, multi-scale expansion convolution of the decoding features is conducive to enhancing the feature representation capability and improving the image resolution; finally, according to the decoded image recognition result, the error source of the encoding and decoding system described in the present invention can be accurately located, and the encoding and decoding system can be corrected.

[0071] 2. After extracting multi-scale features from an image, the present invention uses an attention mechanism to weight features based on channels and spaces, which is conducive to obtaining significant output of the image; the weighted multi-scale features are unified in resolution, which, on the one hand, removes redundant information and noise interference of the image, and on the other hand, facilitates subsequent encoding processing based on multi-scale features, avoiding the compatibility problem of inconsistent resolution. The present invention uses an extreme learning machine model to predict the existence of a target and gives the probability of the existence of a target, which is conducive to quantitatively describing the prediction results; positive and negative samples are set for multiple classifiers, and a reliable support vector machine model is obtained based on staged training of hyperplanes; when the predicted probability of the existence of a target is greater than a threshold, the corresponding multi-scale features are selected and feature fusion is used, and the feature expression ability is improved by enhancing parameters of detail features, which is conducive to improving the encoding quality based on significant features and reducing the interference of image noise; the encoded features are decoded after multi-scale dilation convolution, which is conducive to obtaining more decoding information and improving the ability to describe the detailed features of the decoded image. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 A schematic diagram of a flow chart of a multi-resolution output image encoding and decoding method provided by an embodiment of the present invention;

[0073] Figure 2 A schematic diagram of a flow chart of correcting a coding and decoding system based on a recognition result provided by an embodiment of the present invention;

[0074] Figure 3 A schematic diagram of an output result after decoding processing of an image containing a large-granularity target in an embodiment of the present invention in which inappropriate multi-scale features are selected;

[0075] Figure 4 A schematic diagram of an output result after decoding processing of an image containing a large-granularity target according to an embodiment of the present invention;

[0076] Figure 5 A schematic diagram of the output result after decoding an image containing a small-granularity target according to an embodiment of the present invention;

[0077] Figure 6 A schematic diagram of an output result after decoding processing of an image containing a small-granularity target after increasing detail feature enhancement parameters according to an embodiment of the present invention. DETAILED DESCRIPTION

[0078] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention.

[0079] Embodiment 1:

[0080] A method for encoding and decoding images with multiple resolution outputs, such as Figure 1 As shown, the following steps are included:

[0081] S1: Extract multi-scale features of the image and build a multi-scale feature library:

[0082] S11: Use multiple convolutional networks to extract multi-scale features from images:

[0083] ;

[0084] in, represents the total number of scales, Indicates the scale number, represents the image to be extracted with multi-scale features, Represents the image In the Convolution operations are performed at different scales. Representing images In the Multi-scale features at different scales;

[0085] S12: In multi-scale features Apply the attention mechanism to generate weighted features :

[0086] S121: Based on the channel attention mechanism, multi-scale features Perform channel weighting to obtain channel-weighted multi-scale features:

[0087] ;

[0088] in, represents the multi-scale features after channel weighting, represents channel-wise dot product, represents the channel weight; represents the activation function, represents the fully connected layer, represents the global average pooling operation;

[0089] S122: Based on the spatial attention mechanism, multi-scale features After spatial weighting, the spatially weighted multi-scale features are:

[0090] ;

[0091] in, represents the multi-scale features after spatial weighting, represents the spatial weight; represents the convolution operation, Express Perform maximum pooling, Express Perform average pooling, Indicates channel splicing, Indicates channel concatenation of maximum pooling and average pooling;

[0092] S123: Based on the channel-weighted multi-scale features and the spatial-weighted multi-scale features, a weighted multi-scale feature is generated:

[0093] ;

[0094] S13: The highest resolution value among all weighted multi-scale features is taken as the target resolution ,in represents the target height resolution, Represents the target width resolution; for each weighted multi-scale feature , calculate the scaling ratio of its resolution to the target resolution:

[0095] ;

[0096] in, Indicates the scaling factor of the height resolution, Indicates the scaling factor of the width resolution, express High resolution, express Width resolution;

[0097] S14: Weighted multi-scale features , by periodically arranging and expanding the resolution, the multi-scale features after the expanded resolution are obtained and a multi-scale feature library is formed:

[0098] ;

[0099] in, Represents the multi-scale features after the expanded resolution, Representing multi-scale features Index values ​​in height and width direction, Indicates rounding down; It is a multi-scale feature library.

[0100] S2: Based on the multi-scale feature library, the extreme learning machine model and the support vector machine model are trained to predict the target existence probability and target size:

[0101] S21: The multi-scale feature library obtained in step S1 Input the trained extreme learning machine model , predict the probability of target existence:

[0102] Will Input the trained extreme learning machine model , Output decision function ;

[0103] if , the predicted target does not exist, the predicted target exists probability ;

[0104] if , predicting the existence of the target, the predicted probability of the target existence is expressed as:

[0105] ;

[0106] in, represents the predicted probability of the target existence, represents an exponential function, the probability of the target existence satisfies ；

[0107] S22: The multi-scale feature library obtained in step S1 Input the trained support vector machine model , predict the target size;

[0108] pass classifiers to identify the target size category and calculate all The scores of the classifiers are calculated, and the target size category corresponding to the classifier with the highest score is selected as the predicted target size.

[0109] In the embodiment of the present invention, the support vector machine model The training process is as follows:

[0110] Multi-scale feature library Add tags :

[0111] ;

[0112] in, Represents the total number of target size categories, express The Elements :

[0113] ;

[0114] Construct training data based on labeled multi-scale feature library, the number of classifiers and the total number of target size categories Consistent, use the training data to train all classifiers in sequence;

[0115] Use the training data to train classifiers and obtain optimized classifier parameters:

[0116] ;

[0117] in, Indicates The weight vector of the classifiers is used to define the classification hyperplane, Indicates The bias term of the classifier, Indicates The slack variables of the classifier are used to define the misclassified samples. Indicates The regularization parameter of the classifier is used to balance the complexity of the classification boundary and the tolerance of misclassified samples, and satisfies , Express Taking the second-order norm, Indicates taking the minimum value, and the obtained Parameters are the optimized classifier parameters;

[0118] Label The Elements The multi-scale feature library is marked as a positive sample, label The Elements The multi-scale feature library is marked as negative samples, and the constraints are set as:

[0119] ;

[0120] in, is the decomposition matrix of the kernel function, the kernel function , Represents the mapping function between the hyperplane and the classification accuracy.

[0121] S3: Set the target existence probability threshold based on experience or experimental data; when the predicted target existence probability is greater than the target existence probability threshold In this case, based on the predicted target size, the multi-scale features obtained in step S1 are Select multiple scale features that match the predicted target size and encode them:

[0122] S31: Represents the selected multi-scale features, Extract detail features:

[0123] ;

[0124] in, Express Extracted detail features, represents the detail feature enhancement parameter, Express Features obtained after smoothing; represents a two-dimensional Gaussian filter kernel, Represents the convolution operation;

[0125] S32: Detail features Multi-scale features with selection Fusion, get the encoding features:

[0126] ;

[0127] in, represents the encoding feature, Indicates that and Make stitching.

[0128] In the embodiment of the present invention, the detail feature enhancement parameter When , it means that the detail features account for a large proportion, which is conducive to identifying small-grained targets; the detail feature enhancement parameter When , it indicates that the global features account for a large proportion, which is conducive to identifying large-grained targets.

[0129] In the embodiment of the present invention, if the predicted target existence probability is not greater than the threshold, the process is terminated and no further processing is performed.

[0130] S4: At the receiving end, the encoded features obtained in step S3 are subjected to multi-scale expansion convolution and decoding processing:

[0131] S41: Perform multi-scale dilated convolution on the encoded features to obtain a dilated matrix group:

[0132] ;

[0133] in, represents the elements of the convolution kernel, represents the element value in the encoded feature, Indicates that the index in the encoded feature is Elements of represents the expansion rate, Take separately , is the expansion matrix, is the expansion matrix group;

[0134] S42: Expanded matrix group Performing decoding processing to obtain a decoded image;

[0135] ;

[0136] in, represents the decoded image obtained, Represents the expansion matrix Bilinear interpolation is used for upsampling. Indicates that and Perform feature fusion, It means that convolution is used to refine the fused features, reduce the number of channels and enhance the representation ability.

[0137] S5: Based on the decoded image obtained in step S4, the target is identified, and the system is corrected based on the identification result:

[0138] Based on the decoded image obtained in step S4, target recognition is performed;

[0139] If the target does not exist, it means that the threshold in step S3 The setting is unreasonable, update the threshold ,in, Indicates the threshold update interval, jumps to step S3 to re-encode;

[0140] If the target exists, but the target recognition size category is inconsistent with the target real size category, it means that the detail feature enhancement parameter in step S31 is Improper settings, increase The value of jumps to step S3 for re-encoding;

[0141] If the target recognition accuracy is lower than 60%, it means that the expansion rate in step S41 is If the value is inappropriate, jump to step S41 and update the expansion rate , recalculate the expansion matrix group.

[0142] In the embodiment of the present invention, step S1 obtains a multi-resolution output of the image by extracting multi-scale features of the image; the resolution of the multi-scale features is expanded, and the feature resolutions of different scales are consistent, but the amount of information is different; the large-scale features have less information, and the small-scale features have more information, and there is a non-overlapping part in the information amount of the large-scale features and the small-scale features, and subsequently the appropriate multi-scale features are selected for processing according to the characteristics of the image.

[0143] Step S2 predicts the target size using a scale feature library with a large amount of information, ensuring the output accuracy of the support vector machine model; Step S3 selects appropriate scale features from the size feature library based on the predicted target size, and only selects the most matching feature information for subsequent processing, reducing the workload of the encoding stage and improving encoding efficiency; Step S4 combines the multi-scale dilated convolution method with the bilinear interpolation method to improve the accuracy of detail fitting on the convolution layer.

[0144] Step S5 processes the target recognition result according to different situations: if the target does not exist, it means that step S3 is based on the threshold value. If there is a problem with the secondary judgment of whether the target exists or not, after adjusting the threshold, continue to execute step S3 and subsequent operations; if the target exists but the target recognition result is wrong, it means that there is a problem with the downsampling encoding in step S3, the obtained downsampling features are not prominent or the target features are not sensitive, resulting in the final decoded image being unable to recognize the target, and it is necessary to adjust the detail feature enhancement parameters in step S31, change the feature distribution in the encoded image, increase the proportion of detail features, and then continue to execute step S3 and subsequent operations; when the target recognition accuracy rate obtained after processing a large number of images using the system described in the present invention is low, it means that the target recognition is correct after some images are processed, and the threshold and detail feature enhancement parameters in step S3 are reasonably set. This situation is analyzed and corrected from the decoding stage; the value of the expansion rate in the decoding stage affects the ability to enhance detail features and the decoding effect, so after updating the expansion rate, continue to execute step S4 and subsequent operations.

[0145] To further illustrate, Figure 3-6 As shown, taking an image containing a large-grained target as an example, in step S3, based on the predicted target size, the multi-scale feature library In step S3, appropriate multi-scale features are selected for processing, and finally a decoded image is obtained and target recognition is performed, and the recognition result is correct; when unreasonable multi-scale features are selected for processing in step S3, the target cannot be correctly recognized in the end.

[0146] Taking an image containing a small particle size target as an example, the present invention is used for processing, and the detail feature enhancement parameter When the value is 0.2, the result of target recognition is that the target exists, but the target recognition size category is inconsistent with the target real size category; when identifying small-grained targets, it is required that the detail features account for a large proportion, and the detail feature enhancement parameter , so Set it to 0.8, and re-execute step S3 and subsequent operations; the target recognition result is correct.

[0147] Embodiment 2: The present invention also provides an image encoding and decoding system with multi-resolution output, comprising the following modules:

[0148] Multi-scale feature module: extracts multi-scale features of images and builds a multi-scale feature library;

[0149] Target existence probability and target size prediction module: Based on the multi-scale feature library, an extreme learning machine model and a support vector machine model are trained to predict the target existence probability and target size;

[0150] Encoding module: setting a target existence probability threshold, when the predicted target existence probability is greater than the target existence probability threshold, selecting a multi-scale feature matching the predicted target size in the multi-scale feature library, encoding it, and obtaining the encoded feature;

[0151] Decoding module: The receiving end of the encoding and decoding system receives the encoded features, performs multi-scale dilated convolution on the encoded features, and decodes them to obtain a decoded image;

[0152] System correction module: Identify the target based on the decoded image and correct the encoding and decoding system based on the identification result.

[0153] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method 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, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0154] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0155] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for encoding and decoding an image with multi-resolution output, characterized in that: The method comprises: S1: Extract multi-scale features of the image and build a multi-scale feature library; S11: Use multiple convolutional networks to extract multi-scale features from images: ； in, represents the total number of scales, Indicates the scale number, represents the image to be extracted with multi-scale features, Represents the image In the Convolution operations are performed at different scales. Representing images In the Multi-scale features at different scales; S12: Apply the attention mechanism on the multi-scale features to generate weighted multi-scale features; S13: The highest resolution value among all weighted multi-scale features is taken as the target resolution ,in represents the target height resolution, Indicates the target width resolution; For each weighted multi-scale feature, calculate the scaling ratio of its resolution to the target resolution: ； in, Indicates the scaling factor of the height resolution, Indicates the scaling factor of the width resolution, represents the height resolution of the weighted multi-scale features, Represents the width resolution of weighted multi-scale features; S14: For the weighted multi-scale features, the resolution is expanded by periodic arrangement to obtain multi-scale features with expanded resolution and form a multi-scale feature library: ； in, Represents the multi-scale features after the expanded resolution, Represents the index value of the multi-scale feature in the height and width directions after the resolution is expanded. represents the weighted multi-scale features, Indicates rounding down; It is a multi-scale feature library; S2: Based on the multi-scale feature library, an extreme learning machine model and a support vector machine model are trained to predict the target existence probability and target size; S3: setting a target existence probability threshold, and when the predicted target existence probability is greater than the target existence probability threshold, selecting a multi-scale feature matching the predicted target size from the multi-scale feature library, encoding it, and obtaining the encoded feature; S4: At the receiving end of the encoding and decoding system, the received encoding features are subjected to multi-scale expansion convolution to obtain an expansion matrix group, and the expansion matrix group is subjected to decoding processing to obtain a decoded image; S5: Based on the decoded image, identify the target, and modify the encoding and decoding system based on the identification result.

2. The image encoding and decoding method for multi-resolution output according to claim 1, characterized in that: The S12 comprises the following steps: S121: Based on the channel attention mechanism, channel weighting is performed on the multi-scale features to obtain the channel-weighted multi-scale features: ； in, represents the multi-scale features after channel weighting, represents channel-wise dot product, represents the channel weight; represents the activation function, represents the fully connected layer, represents the global average pooling operation; S122: Based on the spatial attention mechanism, the multi-scale features are spatially weighted to obtain the spatially weighted multi-scale features: ； in, represents the multi-scale features after spatial weighting, represents the spatial weight; represents the convolution operation, Indicates the maximum pooling of multi-scale features. Indicates average pooling of multi-scale features. Indicates channel splicing, Indicates channel concatenation of maximum pooling and average pooling; S123: Generate weighted multi-scale features based on the channel-weighted multi-scale features and the spatial-weighted multi-scale features: .

3. The image encoding and decoding method for multi-resolution output according to claim 1, characterized in that: The S2 includes: S21: Input the multi-scale feature library into the trained extreme learning machine model , predict the probability of target existence; S22: Inputting the multi-scale feature library into the trained support vector machine model , predict the target size; pass classifiers to identify the target size category and calculate all The scores of the classifiers are calculated, and the target size category corresponding to the classifier with the highest score is taken as the predicted target size.

4. The image encoding and decoding method for multi-resolution output according to claim 3, characterized in that: The S21 includes: Input the multi-scale feature library into the trained extreme learning machine model , output decision function ; if , the predicted target does not exist, the predicted target exists probability ; if , predicting the existence of the target, the predicted probability of the target existence is expressed as: ； in, represents the predicted probability of the target existence, represents an exponential function, the probability of the target existence satisfies .

5. The image encoding and decoding method for multi-resolution output according to claim 3, characterized in that: The S3 includes: S31: Setting the target existence probability threshold , when the predicted target existence probability is greater than the target existence probability threshold, a multi-scale feature matching the predicted target size is selected in the multi-scale feature library, and detail features are extracted from the selected multi-scale feature: ； in, represents the selected multi-scale features, Represents the detailed features extracted from the selected multi-scale features, represents the detail feature enhancement parameter, Represents the features obtained after smoothing the selected multi-scale features; represents a two-dimensional Gaussian filter kernel, Represents the convolution operation; S32: Fusion of detail features with selected multi-scale features to obtain encoding features: ； in, represents the encoding feature, Indicates that and Make stitching.

6. The image encoding and decoding method for multi-resolution output according to claim 5, characterized in that: The S4 includes: S41: Perform multi-scale dilation convolution on the coding feature to obtain a dilation matrix group: ； in, represents the elements of the convolution kernel, represents the element value in the encoded feature, Indicates that the index in the encoded feature is Elements of represents the expansion rate, The values ​​are 1 and 2 respectively. is the expansion matrix, is the expansion matrix group; S42: Decoding the expanded matrix group to obtain a decoded image; ； in, represents the decoded image obtained, Represents the expansion matrix Bilinear interpolation is used for upsampling. Indicates that and Perform feature fusion, It means that convolution is used to refine the fused features, reduce the number of channels and enhance the representation ability.

7. The image encoding and decoding method for multi-resolution output according to claim 6, characterized in that: The S5 includes: Identify the target and modify the encoding and decoding system based on the identification results: If the target does not exist, it means the target exists probability threshold If the setting is unreasonable, update the threshold and jump to step S3 for re-encoding. The method for updating the target existence probability threshold is as follows: ； in, Indicates the target existence probability threshold update interval; If the target exists, but the target recognition size category is inconsistent with the target real size category, it means that the detail feature enhancement parameters Improper settings, increase The value of jumps to step S3 for re-encoding; If the target recognition accuracy is less than 60%, it means that the expansion rate If the value is inappropriate, jump to step S41, update the expansion rate, and recalculate the expansion matrix group.

8. An image encoding and decoding system with multi-resolution output, characterized in that: include: Multi-scale feature extraction module: extracts multi-scale features of images and builds a multi-scale feature library; Target existence probability and target size prediction module: Based on the multi-scale feature library, an extreme learning machine model and a support vector machine model are trained to predict the target existence probability and target size; Encoding module: setting a target existence probability threshold, when the predicted target existence probability is greater than the target existence probability threshold, selecting a multi-scale feature matching the predicted target size in the multi-scale feature library, encoding it, and obtaining the encoded feature; Decoding module: The receiving end of the encoding and decoding system receives the encoding features, performs multi-scale dilation convolution and decoding processing on the encoding features, and obtains a decoded image; System correction module: identifies the target based on the decoded image, and corrects the encoding and decoding system based on the recognition result; To implement the image encoding and decoding method with multi-resolution output as described in any one of claims 1-7.

Images