A method and device for compressing and decompressing images of underwater net-working robot

Abstract

This invention discloses a method and device for compressing and decompressing images for underwater net-working robots. First, an underwater image compression model based on frequency-domain cross-scale fusion is constructed, comprising a compression module and a decompression module, specifically comprising two encoders, two quantization modules, two arithmetic encoders, two arithmetic decoders, and two decoders. The first encoder includes a frequency-domain cross-scale fusion module, and both the first encoder and the first decoder include a redundancy removal module. The constructed underwater image compression-decompression model is then trained using an underwater net-working image training set. The trained compression module is then used to compress images captured by the underwater net-working robot to obtain corresponding bitstream data. The trained decompression module is then used to reconstruct the underwater image from the compressed bitstream data. This invention improves feature extraction capabilities by capturing complementary information from different receptive fields, while reducing the impact of redundant information and enhancing image compression performance.

Description

Technical Field

[0001] The present invention belongs to the field of image compression coding, and in particular relates to a method and device for compressing and decompressing images of an underwater netting operation robot. Background Art

[0002] In the deep-sea cage aquaculture industry, the use of underwater net-cleaning robots is replacing traditional manual net cleaning operations for safety and efficiency, and is an industry trend. Operators on the ground remotely control the underwater net-cleaning robots through a visual human-machine interface to complete the cleaning process. Given the limited underwater bandwidth, the underwater images captured by the underwater net-cleaning robots must be compressed at a low bit rate before being transmitted to the ground receiver for reconstruction. High-quality reconstruction is essential for subsequent cleaning tasks.

[0003] Underwater imaging is unique and complex. Due to light scattering and the influence of suspended particles, underwater image clarity is poor. Water disturbances and the motion of underwater net-working robots further degrade the quality of detail in underwater images. Existing image compression networks mostly extract features by stacking multiple layers of convolutional kernels with small receptive fields, making it difficult to account for global correlations between features. This results in low compression efficiency and poor reconstruction quality for underwater images, hindering the implementation of subsequent tasks. Summary of the Invention

[0004] The present invention provides a method and device for compressing and decompressing images of an underwater net-working robot, which can improve compression efficiency and reconstruction quality.

[0005] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0006] A method for compressing and decompressing images of an underwater netting operation robot, comprising:

[0007] Step 1: Construct an underwater image compression-decompression model based on frequency domain cross-scale fusion;

[0008] The underwater image compression-decompression model adopts a scale hyper-prior architecture and includes, from input to output, a first encoder, a first quantization module, a first arithmetic encoder, a first arithmetic decoder, and a first decoder; wherein the first encoder includes a frequency domain cross-scale fusion module, and both the first encoder and the first decoder include a de-redundancy module;

[0009] At the output end of the first encoder, a second encoder, a second quantization module, a second arithmetic encoder, a second arithmetic decoder and a second decoder are sequentially arranged;

[0010] The first encoder, the first quantization module, the first arithmetic encoder, the second encoder, the second quantization module, the second arithmetic encoder, and the second decoder together constitute a compression module of the underwater image compression-decompression model; the first arithmetic decoder, the first decoder, the second arithmetic decoder, and the second decoder together constitute a decompression module of the underwater image compression-decompression model;

[0011] Step 2: Use the underwater net image training set to train the constructed underwater image compression-decompression model;

[0012] Step 3: Use the trained compression module to compress the image collected by the underwater net operation robot to obtain the corresponding bit stream data; use the trained decompression module to reconstruct the underwater image from the compressed bit stream data of the underwater image.

[0013] Furthermore, in the compression module of the underwater image compression-decompression model:

[0014] The first encoder is used to map the underwater image x collected by the underwater net operation robot into a low-dimensional latent space for feature extraction and output the latent representation y;

[0015] The second encoder is used to capture the spatial dependencies between different elements in the latent representation y and output a low-dimensional latent variable ;

[0016] The second quantization module is used to convert the continuous latent variables Perform quantization operations to obtain quantized variables ;

[0017] The second arithmetic encoder is used to calculate the quantized variable The probability distribution of the device is encoded to obtain the second bit stream ;

[0018] The first quantization module is used to quantize the continuous potential representation y to obtain the quantized representation ;

[0019] The second decoder is used to To dynamically estimate the quantized representation Standard deviation , construct a mean of 0 and a standard deviation of Gaussian probability model of;

[0020] The first arithmetic encoder is used to quantize the representation according to a Gaussian probability model. Encode to obtain the first bit stream ; Among them, the first bit stream With the second bitstream Together they constitute the compressed bit stream data of the underwater image;

[0021] and / or,

[0022] In the decompression module of the underwater image compression-decompression model:

[0023] The second arithmetic decoder is configured to perform an arithmetic operation according to the acquired second bit stream. Decoding quantized variables ;

[0024] The second decoder is based on the quantized variable Dynamic Estimation Quantization Representation Standard deviation , construct a mean of 0 and a standard deviation of Gaussian probability model of;

[0025] The first arithmetic decoder converts the acquired first bit stream into Decoding to quantized representation ;

[0026] The first decoder then converts the quantized representation Reconstruct the underwater image.

[0027] Furthermore, the underwater net image training set is derived from real deep-sea cage aquaculture scenes, and underwater net images at different depths, environments, and resolutions are selected.

[0028] Furthermore, the frequency domain cross-scale fusion module is implemented by setting a spatial domain path and a frequency domain path and complementing the information between the two paths:

[0029] First, the input feature map F is proportional to the feature channel dimension. Divided into local feature maps and global feature maps ;

[0030] Then in the spatial domain path, the local feature maps are and global feature maps After convolution processing, the local output feature map is obtained by fusion ;

[0031] In addition, in the frequency domain path, on the one hand, the local feature map Convolution processing, on the other hand, for global feature maps Downsampling, convolution and Fourier transform are performed to the frequency domain for processing, and then the feature maps obtained from the two aspects are fused to obtain the global output feature map ;

[0032] Finally, the local output feature map And the global output feature map Splicing along the feature channel dimension, we get the output feature map M of the frequency domain cross-scale fusion module.

[0033] Furthermore, the steps of the spatial domain path are as follows: the local feature map Through the input channel , the output channel is Ordinary convolution obtains feature maps , the global feature map Through the input channel , the output channel is Ordinary convolution obtains feature maps , you can get the local output feature map :

[0034] ;

[0035] in, Indicates the number of channels of the feature map F before the input frequency domain cross-scale fusion module, Represents the number of channels of the feature map M output from the frequency domain cross-scale fusion module, and are all hyperparameters, is the ratio of local features in the input channel, is the ratio of local output in the output channel.

[0036] Furthermore, the local feature map Through the input channel , the output channel is Ordinary convolution obtains feature maps ;

[0037] The global feature map First, downsample through an average pooling layer, reduce the height and width by half, and then pass through a convolution block including a convolution layer, a BN layer and a ReLU activation function, and then pass through three branches respectively: the first branch first passes the feature map through the fast Fourier transform, extracts the real and imaginary parts of the complex result and splices them, then passes through the convolution layer with a convolution kernel of 1×1, the BN layer and the ReLU activation function, and finally separates the real and imaginary parts again and performs an inverse fast Fourier transform, retaining the real part as the output; the second branch outputs directly without any processing; the third branch first divides the feature map into 4 parts along the height and width respectively and re-splices it into a feature map along the channel dimension, processes it through the Fourier unit module, and then copies four copies for stacking, and finally outputs a feature map with the same height and width as the input branch feature map; then the outputs of the three branches are added and passed through a convolution layer with a convolution kernel of 1×1 to obtain the feature map ; Wherein, the processing method of the first branch is recorded as the Fourier unit module processing;

[0038] Finally, the feature map obtained by the frequency domain path and Add together to get the global output feature map :

[0039] .

[0040] Furthermore, the de-redundancy module is applied after the last convolutional layer of the first encoder and before the first convolutional layer of the first decoder; the de-redundancy module first copies the input feature map, then initializes a 3×3 convolutional layer, applies the convolutional layer to some channels of the input feature map, and keeps the remaining channels unchanged, and performs forward propagation using a slicing method.

[0041] A device for compressing images of an underwater net-working robot, comprising a compression module trained in any of the above-mentioned compression and decompression methods;

[0042] The first encoder is used to map the underwater image x collected by the underwater net operation robot into a low-dimensional latent space for feature extraction and output the latent representation y;

[0043] The second encoder is used to capture the spatial dependencies between different elements in the latent representation y and output a low-dimensional latent variable z;

[0044] The second quantization module is used to quantize the continuous latent variable z to obtain the quantized variable ;

[0045] The second arithmetic encoder is used to calculate the quantized variable The probability distribution of the device is encoded to obtain the second bit stream ;

[0046] The first quantization module is used to quantize the continuous potential representation y to obtain the quantized representation ;

[0047] The second decoder is used to To dynamically estimate the quantized representation Standard deviation , construct a mean of 0 and a standard deviation of Gaussian probability model of;

[0048] The first arithmetic encoder is used to quantize the representation according to a Gaussian probability model. Encode to obtain the first bit stream ; Among them, the first bit stream With the second bitstream Together they constitute the compressed bit stream data of the underwater image.

[0049] A decompression device for underwater net-working robot images, comprising a decompression module trained in any of the above-mentioned compression and decompression methods;

[0050] The second arithmetic decoder is used to obtain the second bit stream Decoding quantized variables ;

[0051] The second decoder is based on the quantization variable Dynamic Estimation Quantization Representation Standard deviation , construct a mean of 0 and a standard deviation of Gaussian probability model of;

[0052] The first arithmetic decoder converts the acquired first bit stream into Decoding to quantized representation ;

[0053] The first decoder then quantizes the Reconstruct the underwater image.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] (1) The computational complexity and parameter count of the frequency domain cross-scale fusion module described in the present invention are comparable to those of ordinary convolution. By complementing the information between the spatial domain path and the frequency domain path, the receptive field is expanded, richer information is transmitted, and the model obtains better compression performance.

[0056] (2) The present invention inserts a de-redundancy module into the first encoder and the first decoder, applies ordinary convolution to some input channels, and keeps the remaining channels unchanged, thereby reducing redundant calculations and memory accesses, and further improving the compression rate and reconstruction effect.

[0057] (3) The present invention is highly practical and can achieve low-bit-rate and high-fidelity compression of underwater net images in different environments, reducing the pressure of underwater transmission and facilitating the implementation of underwater net operation robot tasks and subsequent detection work. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Schematic diagram of the overall framework of an embodiment of the present invention.

[0059] Figure 2 Schematic diagram of a frequency domain cross-scale fusion module according to an embodiment of the present invention.

[0060] Figure 3 Schematic diagram of the compression process of the underwater net operation robot according to an embodiment of the present invention.

[0061] Figure 4 This is a schematic diagram of the ground server decompression process according to an embodiment of the present invention.

[0062] Figure 5 2 is a schematic diagram comparing rate-distortion results according to an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The following is a detailed description of an embodiment of the present invention. This embodiment is based on the technical solution of the present invention, provides a detailed implementation method and a specific operation process, and further explains the technical solution of the present invention.

[0064] Example 1

[0065] This embodiment provides a method for compressing and decompressing images of an underwater net-working robot, comprising the following steps:

[0066] Step 1: Construct an underwater image compression-decompression model based on frequency domain cross-scale fusion.

[0067] The underwater image compression-decompression model adopts a scale super prior architecture, which includes a first encoder, a first quantization module, a first arithmetic encoder, a first arithmetic decoder and a first decoder from input to output. At the output end of the first encoder, a second encoder, a second quantization module, a second arithmetic encoder, a second arithmetic decoder and a second decoder are also provided in sequence. Figure 1 shown.

[0068] 1. Compression module.

[0069] The first encoder, the first quantization module, the first arithmetic encoder, the second encoder, the second quantization module, the second arithmetic encoder, and the second decoder in the underwater image compression-decompression model together constitute a compression module.

[0070] The first encoder maps the underwater image x collected by the underwater net operation robot to a low-dimensional latent space, performs feature extraction and outputs the latent representation y.

[0071] The second encoder captures the spatial dependencies between different elements in the latent representation y and outputs a low-dimensional latent variable .

[0072] The second quantization module converts the continuous latent variable Perform quantization operations to obtain quantized variables .

[0073] The second arithmetic encoder, according to the quantization variable The probability distribution of the device is encoded to obtain the second bit stream .

[0074] The first quantization module quantizes the continuous potential representation y to obtain a quantized representation .

[0075] The second decoder, according to the quantization variable To dynamically estimate the quantized representation The spatial distribution of Standard deviation , reconstructing each point in the potential representation with a mean of 0 and a standard deviation of The Gaussian probability model provides prior information for the first arithmetic encoder and the first arithmetic decoder.

[0076] The first arithmetic encoder quantizes the representation according to the Gaussian probability model Encode to obtain the first bit stream , short sequences are assigned to those with high probability, and long sequences are assigned to those with low probability.

[0077] Among them, the first bit stream obtained by the first arithmetic encoder , and the second arithmetic encoder to obtain the second bit stream , together constitute the compressed bit stream data of the underwater image.

[0078] 2. Decompression module.

[0079] The first arithmetic decoder, the first decoder, the second arithmetic decoder, and the second decoder in the underwater image compression-decompression model together constitute a decompression module.

[0080] The second arithmetic decoder obtains the second bit stream Decoding quantized variables ;

[0081] The second decoder, according to the quantization variable Dynamic Estimation Quantization Representation Standard deviation , construct a mean of 0 and a standard deviation of Gaussian probability model of;

[0082] The first arithmetic decoder converts the acquired first bit stream into Decoding to quantized representation ;

[0083] The first decoder then quantizes the Reconstruct the underwater image.

[0084] 3. Frequency domain cross-scale fusion module.

[0085] The first encoder of the present invention includes a frequency domain cross-scale fusion module, that is, the structure of the first encoder includes a convolution layer A, a nonlinear normalization layer GDN, a frequency domain cross-scale fusion module, GDN, a frequency domain cross-scale fusion module, GDN, a convolution layer B, and a redundancy removal module. The frequency domain cross-scale fusion module is implemented by setting a spatial domain path and a frequency domain path and by complementing the information between the two paths. Figure 2 As shown in Figure 2, the steps of the frequency domain cross-scale fusion module are as follows:

[0086] First, the input feature map F is proportional to the feature channel dimension. Divided into local feature maps and global feature maps :

[0087] ; ; ;

[0088] Among them, H×W represents the spatial resolution of the input feature map, Indicates the number of channels of the input feature map, .

[0089] Then in the spatial domain path, the local feature map Through the input channel , the output channel is And the convolution kernel is 5×5 to obtain the feature map , the global feature map Through the input channel , the output channel is And the convolution kernel is 5×5 to obtain the feature map , you can get the local output feature map :

[0090] ;

[0091] in, Indicates the number of channels of the feature map F before the input frequency domain cross-scale fusion module, Represents the number of channels of the feature map M output from the frequency domain cross-scale fusion module, and are all hyperparameters, is the ratio of local features in the input channel, is the ratio of local output in the output channel, and .

[0092] In addition, in the frequency domain path, on the one hand, the local feature map Convolution processing, on the other hand, for global feature maps Downsampling, convolution and Fourier transform are performed to the frequency domain for processing, and then the feature maps obtained from the two aspects are fused to obtain the global output feature map The steps in the frequency domain path are as follows:

[0093] (1) Local feature map Through the input channel , the output channel is And the convolution kernel is 5×5 to obtain the feature map ;

[0094] (2) The global feature map First, downsample through an average pooling layer, reduce the height and width by half, and then pass through a convolution block including a convolution layer, a BN layer and a ReLU activation function, and then pass through three branches respectively: the first branch first passes the feature map through the fast Fourier transform, extracts the real and imaginary parts of the complex result and splices them, then passes through the convolution layer with a convolution kernel of 1×1, the BN layer and the ReLU activation function, and finally separates the real and imaginary parts again and performs an inverse fast Fourier transform, retaining the real part as the output; the second branch outputs directly without any processing; the third branch first divides the feature map into 4 parts along the height and width respectively and re-splices it into a feature map along the channel dimension, and processes it through a Fourier unit module similar to the first branch, and then copies four copies for stacking, and finally outputs a feature map with the same height and width as the input branch feature map; finally, the outputs of these three branches are added and passed through a convolution layer with a convolution kernel of 1×1 to obtain the feature map .

[0095] (3) The feature map obtained by the frequency domain path and Add together to get the global output feature map :

[0096] .

[0097] Finally, the local output feature map And the global output feature map Splicing along the feature channel dimension, we get the output feature map M of the frequency domain cross-scale fusion module.

[0098] 4. Remove redundant modules.

[0099] The first encoder and the first decoder of the present invention both include a de-redundancy module, which is arranged after the last convolution layer in the first encoder and before the first convolution layer in the first decoder.

[0100] The de-redundancy module first copies the input feature map, then initializes a 3×3 convolution layer, applies the convolution layer to some channels of the input feature map, and keeps the other channels unchanged, and performs forward propagation using a slicing method.

[0101] Step 2: Use the underwater net image training set to train the constructed underwater image compression-decompression model.

[0102] The underwater net image training set described in this embodiment is derived from real deep-sea cage aquaculture scenes photographed by the BlueRov underwater net operation robot on the Mintou No. 1 deep-sea aquaculture platform. Underwater net images at different depths, environments, and resolutions were selected. A total of 2,273 images were collected, of which 2,000 were used as a training set and 273 as a test set, which were used for model training and evaluation respectively.

[0103] This embodiment trains the loss function of the underwater image compression-decompression model to minimize the difference between the reconstructed image and the input image and minimize the bit rate required for encoding, so that the model can achieve optimal performance under the constraint of rate distortion. The reconstruction error is the difference between the reconstructed image and the input image, which is calculated by the mean square error (MSE). The bit rate loss is calculated by the information entropy, which includes the encoding quantization representation. Required number of bits and coded quantization variables The number of bits required. The loss function is:

[0104] ;

[0105] in, represents the loss function, It is a hyperparameter used to balance the compression rate and image reconstruction quality. The smaller the value, the smaller the compressed bitstream file, and the worse the image reconstruction effect after decompression; The larger the value, the larger the compressed bitstream file, and the better the image reconstruction effect after decompression. E represents the reconstruction error. Indicates the compression bit rate, , Representation encoding The number of bits required, Representation encoding The number of bits required.

[0106] In order to enable gradient backpropagation during training, and thus optimize parameters through gradient descent to achieve optimal model performance, the quantization module in the training process uses additive uniform noise instead of a quantizer to implement approximate quantization operations:

[0107] ;

[0108] ;

[0109] in, Indicates the value range is The mean noise, represents the result of approximately quantizing the latent representation y, represents the result of approximately quantizing the latent variable z.

[0110] The quantization module in the test and actual compression process directly uses the quantizer for quantization operations:

[0111] ;

[0112] ;

[0113] in, represents the result of actual quantization y, represents the result of actual quantization z, Indicates quantization.

[0114] Step 3: Use the trained compression module to compress the image collected by the underwater net operation robot to obtain the corresponding compressed bit stream data; use the trained decompression module to reconstruct the underwater image from the compressed bit stream data of the underwater image.

[0115] Example 2

[0116] This embodiment provides an image compression device for an underwater net-working robot. This device is used to compress captured underwater images to obtain a bitstream data file for subsequent storage or transmission. The compression device comprises the trained compression module of the compression and decompression method described in Example 1, and includes a first encoder, a first quantization module, a first arithmetic encoder, a second encoder, a second quantization module, a second arithmetic encoder, and a second decoder.

[0117] The process of using the trained compression module to compress the image collected by the underwater net operation robot in this embodiment is as follows: Figure 3 As shown in the figure: the underwater image x collected by the underwater net operation robot passes through the first encoder to map the underwater image x to a low-dimensional latent space for feature extraction and output the latent representation y; the second encoder captures the spatial dependencies between different elements in the latent representation y and outputs a low-dimensional latent variable z; then the quantizer is used to quantize the continuous latent representation y and the continuous latent variable z respectively to obtain discrete and discrete ; The arithmetic encoder will Convert to bitstream ; Quantify the variable Dynamic estimation by the second decoder Standard deviation , thereby establishing a probability model; the arithmetic encoder converts the discrete Encoded as a bitstream ; Together they constitute compressed bit stream data of underwater images for subsequent storage or transmission.

[0118] Example 3

[0119] This embodiment provides a decompression device for underwater net-working robot images, which is used by a ground server to decode received underwater image bit stream data into a reconstructed image. The device is composed of a decompression module trained in the compression and decompression method described in Example 1, including: a first arithmetic decoder, a first decoder, a second arithmetic decoder, and a second decoder.

[0120] The second arithmetic decoder is used to obtain the second bit stream Decoding quantized variables ;

[0121] The second decoder is based on the quantization variable Dynamic Estimation Quantization Representation Standard deviation , construct a mean of 0 and a standard deviation of Gaussian probability model of;

[0122] The first arithmetic decoder converts the acquired first bit stream into Decoding to quantized representation ;

[0123] The first decoder then quantizes the Reconstruct the underwater image.

[0124] The decompression process is as follows Figure 4 As shown: The bit stream file is decoded into discrete , the quantitative variables Dynamic estimation by the second decoder Standard deviation , thereby reconstructing the probability model, the arithmetic decoder converts the bit stream into Decoding to quantized representation , the discrete The input image is reconstructed by the first decoder.

[0125] The effect of the present invention is tested, and the results are as follows Figure 5As shown in the figure, the vertical axis represents the peak signal-to-noise ratio (PSNR), and the horizontal axis represents the bitrate (bpp). This figure compares the compression performance with that of JPEG2000. A higher PSNR on the vertical axis indicates better reconstruction performance, while a lower bitrate (bpp) on the horizontal axis indicates better compression efficiency. This shows that the method proposed in this invention can achieve the best reconstruction performance at the lowest bitrate.

[0126] The above embodiments are preferred embodiments of the present application. Ordinary technicians in this field can also make various changes or improvements on this basis. Without departing from the overall concept of the present application, these changes or improvements should fall within the scope of protection required by the present application.

Claims

1. A method for compressing and decompressing images of an underwater netting robot, characterized in that: include: Step 1: Construct an underwater image compression-decompression model based on frequency domain cross-scale fusion; The underwater image compression-decompression model adopts a scale hyper-prior architecture and includes, from input to output, a first encoder, a first quantization module, a first arithmetic encoder, a first arithmetic decoder, and a first decoder; wherein the first encoder includes a frequency domain cross-scale fusion module, and both the first encoder and the first decoder include a de-redundancy module; At the output end of the first encoder, a second encoder, a second quantization module, a second arithmetic encoder, a second arithmetic decoder and a second decoder are sequentially arranged; The first encoder, the first quantization module, the first arithmetic encoder, the second encoder, the second quantization module, the second arithmetic encoder, and the second decoder together constitute a compression module of the underwater image compression-decompression model; The first arithmetic decoder, the first decoder, the second arithmetic decoder, and the second decoder together constitute a decompression module of the underwater image compression-decompression model; Step 2: Use the underwater net image training set to train the constructed underwater image compression-decompression model; Step 3: Using the trained compression module, compress the image collected by the underwater net operation robot to obtain corresponding bit stream data; using the trained decompression module, reconstruct the underwater image from the compressed bit stream data; The frequency domain cross-scale fusion module is implemented by setting a spatial domain path and a frequency domain path and complementing the information between the two paths: First, the input feature map F is proportional to the feature channel dimension. Divided into local feature maps and global feature maps ; Then in the spatial domain path, the local feature maps are and global feature maps After convolution processing, the local output feature map is obtained by fusion ; In addition, there is also a frequency domain path: (1) Local feature map Through the input channel , the output channel is Ordinary convolution obtains feature maps ;in, Indicates the number of channels of the feature map F before the input frequency domain cross-scale fusion module, Represents the number of channels of the feature map M output from the frequency domain cross-scale fusion module, and are all hyperparameters, is the ratio of local features in the input channel, is the ratio of local output in the output channel; (2) The global feature map First, downsample through an average pooling layer, reduce the height and width by half, and then pass through a convolution block including a convolution layer, a BN layer and a ReLU activation function, and then pass through three branches respectively: the first branch first passes the feature map through the fast Fourier transform, extracts the real and imaginary parts of the complex result and splices them, then passes through the convolution layer with a convolution kernel of 1×1, the BN layer and the ReLU activation function, and finally separates the real and imaginary parts again and performs an inverse fast Fourier transform, retaining the real part as the output; the second branch outputs directly without any processing; the third branch first divides the feature map into 4 parts along the height and width respectively and re-splices it into a feature map along the channel dimension, processes it through the Fourier unit module, and then copies four copies for stacking, and finally outputs a feature map with the same height and width as the feature map input to the branch; then the outputs of the three branches are added and passed through a convolution layer with a convolution kernel of 1×1 to obtain the feature map ; Wherein, the processing method of the first branch is recorded as the Fourier unit module processing; (3) The feature map obtained by the frequency domain path and Add together to get the global output feature map : ； Finally, the local output feature map And the global output feature map Splicing along the feature channel dimension, we get the output feature map M of the frequency domain cross-scale fusion module.

2. The method for compressing and decompressing images of an underwater netting robot according to claim 1, characterized in that: In the compression module of the underwater image compression-decompression model: The first encoder is used to map the underwater image x collected by the underwater net operation robot into a low-dimensional latent space for feature extraction and output the latent representation y; The second encoder is used to capture the spatial dependencies between different elements in the latent representation y and output a low-dimensional latent variable ; The second quantization module is used to convert the continuous latent variables Perform quantization operations to obtain quantized variables ; The second arithmetic encoder is used to calculate the quantized variable The probability distribution of the device is encoded to obtain the second bit stream ; The first quantization module is used to quantize the continuous potential representation y to obtain the quantized representation ; The second decoder is used to To dynamically estimate the quantized representation Standard deviation , construct a mean of 0 and a standard deviation of Gaussian probability model of; The first arithmetic encoder is used to quantize the representation according to a Gaussian probability model. Encode to obtain the first bit stream ; Among them, the first bit stream With the second bitstream Together they constitute the compressed bit stream data of the underwater image; and / or, In the decompression module of the underwater image compression-decompression model: The second arithmetic decoder is configured to perform an arithmetic operation according to the acquired second bit stream. Decoding quantized variables ; The second decoder is based on the quantized variable Dynamic Estimation Quantization Representation Standard deviation , construct a mean of 0 and a standard deviation of Gaussian probability model of; The first arithmetic decoder converts the acquired first bit stream into Decoding to quantized representation ; The first decoder then converts the quantized representation Reconstruct the underwater image.

3. The method for compressing and decompressing images of an underwater netting robot according to claim 1, characterized in that: The underwater net image training set is derived from real deep-sea cage aquaculture scenes, and underwater net images at different depths, environments, and resolutions are selected.

4. The method for compressing and decompressing images of an underwater netting robot according to claim 1, characterized in that: The steps of the spatial domain path are as follows: transform the local feature map Through the input channel , the output channel is Ordinary convolution obtains feature maps , the global feature map Through the input channel , the output channel is Ordinary convolution obtains feature maps , you can get the local output feature map : ； in, Indicates the number of channels of the feature map F before the input frequency domain cross-scale fusion module, Represents the number of channels of the feature map M output from the frequency domain cross-scale fusion module, and are all hyperparameters, is the ratio of local features in the input channel, is the ratio of local output in the output channel.

5. The method for compressing and decompressing images of an underwater netting robot according to claim 1, characterized in that: The de-redundancy module is applied after the last convolutional layer of the first encoder and before the first convolutional layer of the first decoder. The de-redundancy module first copies the input feature map, then initializes a 3×3 convolutional layer, applies the convolutional layer to some channels of the input feature map, and keeps the remaining channels unchanged, and performs forward propagation using a slicing method.

6. A device for compressing images of an underwater netting robot, characterized in that: Composed of a compression module trained in the compression and decompression method according to any one of claims 1 to 5; The first encoder is used to map the underwater image x collected by the underwater net operation robot into a low-dimensional latent space for feature extraction and output the latent representation y; The second encoder is used to capture the spatial dependencies between different elements in the latent representation y and output a low-dimensional latent variable z; The second quantization module is used to quantize the continuous latent variable z to obtain the quantized variable ; The second arithmetic encoder is used to calculate the quantized variable The probability distribution of the device is encoded to obtain the second bit stream ; The first quantization module is used to quantize the continuous potential representation y to obtain the quantized representation ; The second decoder is used to To dynamically estimate the quantized representation Standard deviation , construct a mean of 0 and a standard deviation of Gaussian probability model of; The first arithmetic encoder is used to quantize the representation according to a Gaussian probability model. Encode to obtain the first bit stream ; Among them, the first bit stream With the second bitstream Together they constitute the compressed bit stream data of the underwater image.

7. A decompression device for underwater netting robot images, characterized in that: Composed of a decompression module trained in the compression and decompression method according to any one of claims 1 to 5; The second arithmetic decoder is used to obtain the second bit stream Decoding quantized variables ; The second decoder is based on the quantization variable Dynamic Estimation Quantization Representation Standard deviation , construct a mean of 0 and a standard deviation of Gaussian probability model of; The first arithmetic decoder converts the acquired first bit stream into Decoding to quantized representation ; The first decoder then quantizes the Reconstruct the underwater image.

Images