A ridge segmentation method based on encoder-decoder architecture combined with strip pooling and ASPP
By using the encoder-decoder architecture in field ridge segmentation combined with belt pooling and ASPP module, the problem of limited information accuracy in the field ridge in the existing technology is solved, and efficient segmentation of narrow and irregular characteristics of field ridge shape is achieved. Improve the accuracy and connectivity of field ridge segmentation.
Patent Information
- Application Number
- CN202410104596.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-25
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-01-25
AI Technical Summary
The prior art is difficult to effectively extract the spatial and texture features of high-resolution images in the segmentation of farmland ridges, resulting in limited information accuracy of field ridges, especially the narrow shape of the field ridges, complex edges and irregular features that are difficult to accurately segment.
The ridge segmentation method based on the encoder-decoder architecture combined with the band pooling module and the ASPP module is adopted. Through layer-by-layer feature extraction, spatial information correlation and feature fusion are enhanced, and the band pooling module is used to strengthen the vertical and horizontal distant directions. Distance dependency, combined with the ASPP module to capture feature information in different receptive fields.
It significantly improves the accuracy and connectivity of field ridge segmentation, enhances the detection effect of narrow and irregular characteristics of field ridge shape, and improves the segmentation accuracy of farmland ridge information.
Smart Images

Figure CN117975004B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of image segmentation, and in particular relates to a field ridge segmentation method based on an encoder-decoder architecture combined with a strip pooling module and an ASPP module. Background Art
[0002] Mapping and monitoring farmland information is one of the most basic uses of remote sensing data in farmland management. In the process of farmland breeding, ridges divide farmland into different crop areas, helping farmers to make reasonable planting arrangements and crop management. Therefore, accurately and efficiently capturing farmland ridge information from low-altitude remote sensing data is of great significance to farmland management and decision-making.
[0003] In recent years, more and more studies have used remote sensing images to analyze farmland distribution information. In the processing of remote sensing images, the method of manually drawing farmland distribution information is easily affected by subjective factors, and the data sources are scattered, which is not conducive to unified information collection and management, and it is difficult to meet the needs of efficient farmland management. The emergence of machine learning provides corresponding algorithm support for the processing of remote sensing images, helping to realize the automatic segmentation of farmland information. Traditional machine learning methods mainly use the relationship between pixels to achieve segmentation, such as using the decision tree classification algorithm to generate a farmland spatial distribution map based on the spectral similarity between image pixels; through the idea of regional division, farmland segmentation is performed according to the average local variance; the original color image is converted into a grayscale image, and the optimal threshold is obtained by minimizing the intuitive fuzzy divergence to detect and segment farmland.
[0004] The implementation of the above methods relies more on the similarity of pixels, lacks the extraction of spatial and texture features of high-resolution images, resulting in limited accuracy of the acquired farmland information. With the rapid development of deep learning, convolutional neural networks have been able to extract rich semantic information, thereby alleviating the above shortcomings. The neural network based on the encoder-decoder structure can perform feature extraction in the encoder, upsample the extracted deep features in the decoder and combine them with shallow features, which can improve the detection effect of edge areas in farmland segmentation tasks. However, in the process of restoring the segmentation results, the information between different receptive fields is lacking. In the study of farmland segmentation, Du et al. used the DeepLabv3+ model with dilated convolution to extract and map crop distribution in order to accurately describe small and irregular fields in farmland. Nevertheless, in the implementation of farmland ridge segmentation tasks, the unique characteristics of ridges with large spans and narrow shapes need to be considered. Summary of the invention
[0005] Aiming at the characteristics of narrow ridge shape, wide receptive field range, complex and irregular edges involved in ridge information in remote sensing images, the present invention proposes a ridge segmentation method based on an encoder-decoder architecture combined with a strip pooling module and an ASPP module, which can accurately segment the ridge information.
[0006] Technical solution:
[0007] A method for segmenting ridges based on an encoder-decoder architecture combined with a strip pooling module and an ASPP module comprises the following steps:
[0008] Step 1: Collect the farmland ridge data set, mark the ridge information in the image, and create a data set;
[0009] Step 2, establishing a segmentation model, using the data set prepared in step 1 to train the field segmentation model, wherein the segmentation model includes an encoder structure for layer-by-layer feature extraction, an ASPP module for enhancing spatial information correlation, a band pooling module for strengthening feature dependency in the band direction, and a decoder structure for fusing deep features with shallow features;
[0010] Step 3: Use the segmentation model trained in step 2 to input the farmland image, and output the farmland ridges and background areas through model prediction.
[0011] Preferably, step 1 includes the following steps: step 1.1, labeling the data set, using ENVI software to label the ridge information in the farmland image, marking the ridge information in the farmland image and other backgrounds with different colors as segmentation data labels; step 1.2, image cropping: uniformly cropping the farmland image and the marked data labels to a resolution of 512*512, and renaming each cropped data, so that the farmland image corresponds to the data label name; step 1.3, sorting the processed ridge data set and dividing it into a training set and a validation set in proportion.
[0012] Preferably, in step 2, the encoder performs layer-by-layer feature downsampling to generate feature maps of each layer, the feature map of the last layer is input into the input end of the ASPP module to capture feature information under different receptive fields, and the output end of the ASPP module is connected to the input end of the decoder for feature upsampling, wherein the decoder module also includes a plurality of band pooling modules connected step by step, and the feature maps of different scales output by the encoder module are respectively connected to the input end of the band pooling module of the corresponding level through a convolution module, and are concat-connected with the output of the upsampling of the previous level as the input of the upsampling of the next level, and this upsampling operation is repeated several times, and the output of the band pooling module of the last level is subjected to convolution processing to form the final result of the decoder.
[0013] Preferably, the encoder downsampling in step 2 comprises the following steps:
[0014] Step 2.1, pass the image to be segmented into the encoder. The input is a farmland image of H*W*3, that is, a two-dimensional image with H*W pixels and 3 channels. The ridge features are preliminarily processed through the convolution block. The convolution block contains two sets of convolution layers, a normalization layer, and an activation layer. The convolution kernel used in the convolution layer is 3*3, the step size is 1, the number of channels is 32, and the features of size H*W*32 are extracted;
[0015] In step 2.2, the features obtained in step 2.1 are passed to the downsampling module in sequence, and downsampling is performed layer by layer. The output of the previous downsampling layer is used as the input of the next downsampling layer. Each downsampling layer obtains a feature map, and the feature map obtained in the last layer is the total output of the encoder module.
[0016] Preferably, the number of downsampling modules is 4, the input pixels are H*W, and the number of channels is C. The significant features are extracted by maximum pooling with a step size of 2. The first layer of downsampling obtains a feature map of H / 2*W / 2 and C channels, and the feature map is input into the convolution block. The second layer of downsampling obtains a feature map of H / 2*W / 2 and 2C channels, and the feature map is input into the convolution block. The third layer of downsampling obtains a feature map of H / 2*W / 2 and 4C channels, and the feature map is input into the convolution block. The last layer of downsampling obtains a feature map of H / 2*W / 2 and 8C channels.
[0017] Preferably, in step 2, the following steps are performed in the dilated convolution ASPP module:
[0018] Step 2.3: Pass the feature map of the last layer of the encoder into four convolution blocks with different dilation rates for parallel processing. The first block is a 1×1 ordinary convolution branch with a dilation rate of 1, and the remaining three blocks are 3×3 convolution block branches with dilation rates of 6, 12, and 18, respectively. The step size is 1, and the number of output channels is half of the number of input channels, that is, the output of each dilation convolution block is H*W*C / 2.
[0019] In step 2.4, the four H*W*C / 2 feature maps obtained in step 2.3 are concat-joined. The number of channels after concatenation is 2C, and the convolution block operation is performed on the concatenated feature map, that is, the image is convolved Conv, normalized BatchNorm, and activated by ReLU in sequence to realize the processing of the concatenated features and restore the number of channels to C / 2 through the above convolution.
[0020] Preferably, in step 2, the decoder process corresponds to the encoder process, and the band pooling module is integrated in the decoder upsampling process, specifically including the following steps:
[0021] Step 2.5, upsample the features outputted in step 2.4, concatenate the upsampled features with the feature maps of the corresponding size outputted by the encoder, and perform convolution operation on the concatenated features;
[0022] Step 2.6, the obtained new feature map is passed to the next level strip pooling module SPBlock;
[0023] Step 2.7, repeat steps 2.5-2.6 to perform step-by-step upsampling and band pooling operations. The feature map output by the last band pooling module is the total output of the decoder module. A convolution operation is performed on the total output to output the final feature result.
[0024] Preferably, in step 2, the strip pooling module process is specifically as follows:
[0025] Step 2.61, use the H*1 strip pooling kernel to move horizontally on the input features, take the average of all feature values in each column, then perform convolution operation with a 1*3 strip convolution kernel and normalize, then copy the features to expand them to H*W size;
[0026] Step 2.62, corresponding to step 2.61, uses a 1*W strip pooling kernel and a 3*1 band convolution kernel for operation. The outputs of the vertical strip pooling and horizontal strip pooling are as follows:
[0027]
[0028] Where x is the input feature map;
[0029] Step 2.63, perform pixel-by-pixel sum operation on the features corresponding to the same position in the feature maps obtained in step 2.61 and step 2.62 to obtain a feature map of size H*W;
[0030] Step 2.64, perform feature processing on the feature map obtained in 2.63 through a convolution operation, then use the sigmoid activation function to map the feature value to between (0, 1), and multiply it with the corresponding pixel of the original input feature map to obtain the output result of the module.
[0031] Preferably, the segmentation of step 3 comprises the following steps:
[0032] Step 3.1, cut the input farmland image into n*m small blocks with a resolution of 512*512;
[0033] Step 3.2, each small image block is passed into the segmentation model trained in step 2 to perform ridge segmentation;
[0034] Step 3.3, concatenate the n*m small-block segmentation results into a large-area farmland segmentation result.
[0035] The present invention also discloses a field ridge segmentation system based on an encoder-decoder architecture combined with strip pooling and ASPP, comprising an encoder module, comprising four groups of down-sampling modules connected in sequence, the down-sampling modules performing step-by-step feature extraction, and the output end of the encoder module being connected to the input end of the ASPP module; the ASPP module comprising a hole convolution module and a Concat splicing module, the hole convolution module performing four convolution operations with different hole rates on the input features in parallel, and outputting four groups of feature maps to the input end of the Concat splicing module respectively, and the output end of the ASPP module being connected to the input end of the decoder module; the decoder module comprising four groups of up-sampling modules, Concat splicing modules and strip pooling modules connected in sequence, the output end of the up-sampling module being connected to the output end of the down-sampling module of corresponding size through the Concat splicing module, and then being connected to the input end of the strip pooling module; the output end of the strip pooling module being connected to the input end of the next-level up-sampling module, and performing the next-level up-sampling; the output end of the last-level strip pooling module being used as the output end of the decoder, and the output end of the decoder performing a convolution and outputting the final segmentation result.
[0036] Compared with the prior art, the present invention has the following beneficial effects:
[0037] The encoder-decoder structure is used as the main architecture in the present invention. The U-shaped structure can effectively retain the shallow edge texture features in the process of feature fusion. According to the long and narrow characteristics of the ridge, a strip pooling module is added to the decoder to strengthen the long-distance dependency of the features in the vertical and horizontal directions. Since the receptive field range of the ridge information in the remote sensing image is wide, the ASPP module is added to the model to enhance the correlation of the global spatial information, thereby improving the connectivity of the ridge segmentation effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 It is a structural diagram of a model in an embodiment of the present invention;
[0039] Figure 2 This is the complete process and effect of using the segmentation model for large-scale farmland segmentation in an embodiment of the present invention. DETAILED DESCRIPTION
[0040] The scheme in the example of the present invention is described in detail below with reference to the accompanying drawings.
[0041] The present invention proposes a ridge segmentation method based on an encoder-decoder architecture combined with a strip pooling module and an ASPP module. The farmland image is labeled using ENVI; the processed farmland image is input into the segmentation model for training. After the training, the ridge segmentation can be performed. Considering the problem of large coverage area, high image resolution and the need to ensure segmentation accuracy, the farmland image to be segmented is cut into multiple fixed-size image blocks, and these images are input into the segmentation model, and then the output image blocks are spliced to achieve large-area farmland ridge segmentation. The detection model is based on an encoder-decoder architecture as a whole. The encoder part extracts high-level features of the image by gradually reducing the image resolution, while the decoder part restores the feature map to the resolution size of the original image through upsampling operations. In order to avoid omissions in ridge feature extraction, ASPP is added to the highest semantic information layer extracted by the encoder to obtain spatial context information at different scales of the feature map. Finally, based on the particularity of small sample feature learning, in the decoder structure, the local features of the encoder output are fused with the features of the upsampled output in sequence, and according to the data characteristics, the fused features are operated with the band spatial attention mechanism to obtain long-distance dependencies in the horizontal and vertical directions, which helps the model to fuse features that are more in line with the ridges during the upsampling process.
[0042] like Figure 1 As shown, the following steps are included:
[0043] Step 1: Create a farmland ridge segmentation dataset by annotating the collected farmland images. When training the segmentation model, compare the ridge area obtained by the training segmentation model with the manually annotated ridge information to test the accuracy of the segmentation model. The specific steps are as follows:
[0044] Step 1.1, collect farmland remote sensing images, annotate each farmland image according to the actual ridge distribution information, and use ENVI software to mark the ridge information and other background in the farmland image with different colors. The annotation process is to use the Region of Interest (ROI) Tool in ENVI software to establish a category region of interest, first name the label, then mark the ridge area on the image and save it.
[0045] Step 1.2: Cut the farmland images and labels into multiple small images with a resolution of 512*512.
[0046] Step 1.3, rename the cropped farmland images and labels so that the images and label names correspond one to one.
[0047] Step 1.4: sort the processed field ridge data set and divide it into training set and validation set in a ratio of 5:1. The training set is used to train the segmentation model, and the validation set is used to test the trained model.
[0048] Step 2: Use the field segmentation dataset in step 1 to train the segmentation model. The segmentation model uses the encoder-decoder structure as the main architecture, including encoder downsampling for feature extraction, and then passing the deepest global features output by the encoder into the dilated convolution ASPP module, and then passing the processed module into the decoder for upsampling and feature fusion operations. Among them, the features of the downsampled output are combined with the features of the encoder upsampled, and shallow texture features are combined while obtaining deep semantic information. After upsampling, the strip pooling module is integrated to perform horizontal and vertical pooling operations on the input features to obtain long-distance dependencies in the form of feature strips.
[0049] Specifically, it includes the encoder process, the intermediate ASPP processing, the decoder process with band pooling, and the field ridge segmentation process using the model.
[0050] Encoder process:
[0051] Step 2.1, pass the processed field segmentation dataset into the segmentation model, and input the farmland image of H*W*3, that is, a two-dimensional image with H*W pixels and 3 channels. The original input image is processed with shallow features through the convolution block. The convolution block contains two sets of convolution Conv, normalized BN and activated ReLU operations, in which the convolution layer uses a convolution operation with a convolution kernel of 3*3, a step size of 1 and a channel number of 32 to extract features of size H*W*32.
[0052] Step 2.2: Input the feature map output from step 2.1 into the downsampling module, that is, the input feature is a feature map of H*W*C, extract the significant features through the maximum pooling with a step size of 2, and output the feature map of H / 2*W / 2*C after pooling. Then input the feature map into the convolution block for feature extraction. The convolution block is the same as the convolution block in step 2.1, but the output channel is twice the input channel, that is, the output is a feature map of H / 2*W / 2*2C, and it is used as the input of the next stage. Repeat in sequence, a total of four groups of downsampling modules, outputting four groups of features of different sizes.
[0053] Intermediate ASPP processing:
[0054] Step 2.3, obtain the last layer of features of the encoder, and pass the last layer of features into the dilated convolution ASPP module. This module first processes the feature map separately in four parallel branches, each branch contains convolution blocks with different dilation rates, and introduces the "expansion rate" parameter based on traditional convolution to control the sampling interval between the convolution kernel and the input feature map. The first branch is a 1*1 ordinary convolution branch with a dilation rate of 1, and the remaining three branches are 3*3 convolution block branches with dilation rates of 6, 12 and 18, with a step size of 1. The output channel of each convolution block is half of the number of input channels, that is, the output of each dilated convolution block is H*W*C / 2;
[0055] In step 2.4, the features of the original input ASPP module are globally averaged and concatenated with the H*W*C / 2 feature maps of the four different dilated convolution outputs obtained in step 2.3, and superimposed according to the number of channels to obtain a new feature map. The feature maps output after the superposition of each ASPP layer are convoluted, normalized, and ReLU is activated to reduce the dimension of the feature map to the given number of out_channel channels, and a simple feature processing is performed through the dropout operation to output the final result.
[0056] Decoder process with band pooling:
[0057] Step 2.5, first upsample the features output in step 2.4, upsample the H*W*C features to 2H*2W*C feature maps, and at the same time, perform a convolution operation on the corresponding 2H*2W*C feature map output by the encoder, output 2H*2W*C features, and then concat the upsampled features with the convolution features to obtain a 2H*2W*2C feature map, and then change the number of feature channels through a convolution block to output a 2H*2W*C / 2 feature map.
[0058] In step 2.6, the feature map output in step 2.5 is passed into the SPBlock strip pooling module, and the horizontal and vertical feature extraction operations are implemented through two parallel branches. Taking one of the branches as an example, a H*1 strip pooling kernel is used to move horizontally, and the feature values in each column are averaged. Then, a 1*3 strip convolution kernel is used to perform a Conv operation and normalize the BN. After that, the feature copy is expanded to a size of H*W*C by copying. Similarly, the other branch uses a 1*W strip pooling and a 3*1 strip convolution kernel to perform the operation, and after copying, a H*W*C size feature map is output. The output of the vertical and horizontal strip pooling mentioned is shown in the formula: Where x is the input feature map. After that, the feature maps obtained by the above horizontal and vertical operations are processed, and the features corresponding to the same pixel position are summed pixel by pixel to obtain a new H*W*C feature map. Next, the convolution operation is used to change the number of feature channels, and the feature value is mapped to (0, 1) through the sigmoid activation function. The mapped value is used as a coefficient to multiply the feature map in the original input SPBlock, and finally the processing result of the SPBlock band pooling module is output.
[0059] Step 2.7, then repeat steps 2.5-2.6 three times to implement upsampling and band pooling operations, and perform output convolution operations on the feature map output by the last layer to output the final feature results.
[0060] Use the model to realize the field ridge segmentation process:
[0061] like Figure 2 As shown in the figure, the trained segmentation model is read, and the farmland image to be segmented is divided into n*m small blocks according to the resolution of 512*512. Then, each small block image obtained by cutting is passed into the previously trained field ridge segmentation model, and the field ridge and background are marked with different colors through model prediction to display the segmentation result. Finally, the predicted n*m small block segmentation results are spliced into the segmentation result of a large area of farmland.
[0062] The present invention also discloses a ridge segmentation system model based on an encoder-decoder architecture combined with band pooling and ASPP, comprising
[0063] The encoder module includes four groups of down-sampling modules connected in sequence, the down-sampling modules perform step-by-step feature extraction, and the output end of the encoder module is connected to the input end of the ASPP module;
[0064] The ASPP module includes a dilated convolution module and a Concat module. The dilated convolution module performs four convolution operations with different dilation rates on the input features in parallel, and outputs four sets of feature maps to the input end of the Concat module respectively. The output end of the ASPP module is connected to the input end of the decoder module.
[0065] The decoder module includes four groups of upsampling modules, Concat splicing modules and strip pooling modules connected in sequence. The output end of the upsampling module is connected to the output end of the downsampling module of the corresponding size through the Concat splicing module, and then connected to the input end of the strip pooling module; the output end of the strip pooling module is connected to the input end of the next level upsampling module to perform the next level upsampling; the output end of the last level of strip pooling module is used as the output end of the decoder, and the output end of the decoder performs a convolution and then outputs the final segmentation result.
[0066] The model disclosed in this application is compared with other conventional segmentation models, as shown in Table 1 below:
[0067] Table 1 Experimental comparison between conventional segmentation model and this model
[0068]
[0069] Experimental comparison shows that the segmentation method and segmentation system model proposed in this application perform best in terms of accuracy, recall and mIoU evaluation indicators.
[0070] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A method for field ridge segmentation based on an encoder-decoder architecture combined with band pooling and ASPP, characterized in that: The following steps are involved: Step 1: Collect the farmland ridge data set, mark the ridge information in the image, and create a data set; Step 2, establish a segmentation model, use the data set prepared in step 1 to train the field segmentation model, wherein the segmentation model includes an encoder structure for layer-by-layer feature extraction, an ASPP module for enhancing spatial information relevance, a strip pooling module for strengthening feature dependencies in the strip direction, and a decoder structure for fusing deep features with shallow features. Specifically, the encoder performs layer-by-layer feature downsampling to generate feature maps of each layer, and the feature map of the last layer is input into the input end of the ASPP module to capture feature information under different receptive fields. The output end of the ASPP module is connected to the input end of the decoder for feature upsampling, wherein the decoder module also includes a number of strip pooling modules connected step by step, and the feature maps of different scales output by the encoder module are respectively connected to the input end of the strip pooling module of the corresponding level through a convolution module, and are concat-connected with the output of the previous level upsampling as the input of the next level upsampling, and this upsampling operation is repeated several times, and the output of the last level of strip pooling module is subjected to convolution processing to form the final result of the decoder. Among them, the following steps are performed in the hole convolution ASPP module: Step 2.3: Pass the feature map of the last layer of the encoder into four convolution blocks with different dilation rates for parallel processing. The first block is a 1×1 ordinary convolution branch with a dilation rate of 1, and the remaining three blocks are 3×3 convolution block branches with dilation rates of 6, 12, and 18, respectively. The step size is 1, and the number of output channels is half of the number of input channels, that is, the output of each dilation convolution block is H*W*C / 2. Step 2.4, concatenate the four H*W*C / 2 feature maps obtained in step 2.3, the number of channels after concatenation is 2C, and perform convolution block operation on the concatenated feature maps, that is, perform convolution Conv, normalization BatchNorm, activation function ReLU operation on the image in sequence and output it, and restore the number of channels to C / 2 through the above convolution; The decoder process corresponds to the encoder process, and the band pooling module is integrated in the decoder upsampling process. Specifically, Step 2.5, upsample the features outputted in step 2.4, concatenate the upsampled features with the feature maps of the corresponding size outputted by the encoder, and perform convolution operation on the concatenated features; Step 2.6, obtain the new feature map and pass it to the next level strip pooling module SPBlock; Step 2.7, repeat steps 2.5-2.6 to perform step-by-step upsampling and band pooling operations. The feature map output by the last band pooling module is the total output of the decoder module. A convolution operation is performed on the total output to output the final feature result. Step 3: Use the segmentation model trained in step 2 to input the farmland image, and output the farmland ridges and background areas through model prediction.
2. The method for ridge segmentation based on an encoder-decoder architecture combined with band pooling and ASPP according to claim 1, characterized in that: Step 1 includes the following steps: Step 1.1, labeling the data set, using ENVI software to label the ridge information in the farmland image, marking the ridge information in the farmland image and other backgrounds with different colors as segmentation data labels; Step 1.2, image cropping: crop the farmland images and marked data labels to a uniform resolution size, and rename each cropped data to correspond the farmland images to the data label names; Step 1.3: sort the processed field ridge data set and divide it into training set and validation set in proportion.
3. The method for field ridge segmentation based on an encoder-decoder architecture combined with band pooling and ASPP according to claim 1, characterized in that: The encoder downsampling in step 2 includes the following steps: Step 2.1, pass the image to be segmented into the encoder. The input is a farmland image of H*W*3, that is, a two-dimensional image with H*W pixels and 3 channels. The ridge features are preliminarily processed through the convolution block. The convolution block contains two sets of convolution layers, a normalization layer, and an activation layer. The convolution kernel used in the convolution layer is 3*3, the step size is 1, the number of channels is 32, and the features of size H*W*32 are extracted; In step 2.2, the features obtained in step 2.1 are passed to the downsampling module in sequence, and downsampling is performed layer by layer. The output of the previous downsampling layer is used as the input of the next downsampling layer. Each downsampling layer obtains a feature map, and the feature map obtained in the last layer is the total output of the encoder module.
4. The method for field ridge segmentation based on an encoder-decoder architecture combined with band pooling and ASPP according to claim 1, characterized in that: The number of downsampling modules is 4, the input pixels are H*W, and the number of channels is C. The feature map is extracted by maximum pooling with a step size of 2. The first layer of downsampling obtains a feature map of H / 2*W / 2 and C channels, and then the feature map is input into the convolution block. The second layer of downsampling obtains a feature map of H / 2*W / 2 and 2C channels, and then the feature map is input into the convolution block. The third layer of downsampling obtains a feature map of H / 2*W / 2 and 4C channels, and then the feature map is input into the convolution block. The last layer of downsampling obtains a feature map of H / 2*W / 2 and 8C channels.
5. The method for field ridge segmentation based on an encoder-decoder architecture combined with band pooling and ASPP according to claim 1, characterized in that: In step 2, the strip pooling module process is as follows: Step 2.61, use the H*1 strip pooling kernel to move horizontally on the input features, take the average of all feature values in each column, then perform convolution operation with a 1*3 strip convolution kernel and normalize, then copy the features to expand them to H*W size; Step 2.62, corresponding to step 2.61, uses a 1*W strip pooling kernel and a 3*1 band convolution kernel for operation. The outputs of the vertical strip pooling and horizontal strip pooling are as follows: , , Where x is the input feature map; Step 2.63, perform pixel-by-pixel sum operation on the features corresponding to the same position in the feature maps obtained in step 2.61 and step 2.62 to obtain a feature map of size H*W; Step 2.64, perform feature processing on the feature map obtained in 2.63 through convolution operation, then use the sigmoid activation function to map the feature value to between (0, 1), and multiply it with the corresponding pixel of the original input feature map to obtain the output result of this module.
6. The method for field ridge segmentation based on an encoder-decoder architecture combined with band pooling and ASPP according to claim 1, characterized in that: The segmentation of step 3 includes the following steps: Step 3.1, cut the input farmland image into n*m small blocks with a resolution of 512*512; Step 3.2, each small image block is passed into the segmentation model trained in step 2 to perform ridge segmentation; Step 3.3, concatenate the n*m small-block segmentation results into a large-area farmland segmentation result.
7. A field ridge segmentation system based on an encoder-decoder architecture combined with strip pooling and ASPP, characterized in that: Based on the field ridge segmentation method as claimed in claim 1, comprising: The encoder module includes four groups of down-sampling modules connected in sequence, the down-sampling modules perform step-by-step feature extraction, and the output end of the encoder module is connected to the input end of the ASPP module; The ASPP module includes a dilated convolution module and a Concat module. The dilated convolution module performs four convolution operations with different dilation rates on the input features in parallel, and outputs four sets of feature maps to the input end of the Concat module respectively. The output end of the ASPP module is connected to the input end of the decoder module. The decoder module includes four groups of upsampling modules, Concat splicing modules and strip pooling modules connected in sequence. The output end of the upsampling module is connected to the output end of the downsampling module of the corresponding size through the Concat splicing module, and then connected to the input end of the strip pooling module; the output end of the strip pooling module is connected to the input end of the next level upsampling module to perform the next level upsampling; the output end of the last level of strip pooling module is used as the output end of the decoder, and the output end of the decoder performs a convolution and then outputs the final segmentation result.
