A remote sensing image segmentation method and device based on an improved ERFNet network

Abstract

The application discloses a kind of based on improved ERFNet network's remote sensing image segmentation method and device, it is related to image semantic segmentation technical field;Through obtaining remote sensing image;Remote sensing image is input into the improved ERFNet semantic segmentation network model after training, and remote sensing image segmentation result is obtained, wherein: the improved ERFNet semantic segmentation network model includes the encoder module of adding attention pyramid pooling module and the decoder module of using light flow guide alignment module to replace deconvolution;Compared with traditional ERFNet network, the application increases attention pyramid pooling module in the encoder, uses light flow guide alignment module to replace deconvolution in decoder, fully fuses multi-scale information, ensures image restoration accuracy, effectively improves the segmentation precision of small object in remote sensing high-resolution image and overall image.

Description

TECHNICAL FIELD

[0001] The application relates to a remote sensing image segmentation method and device based on an improved ERFNet network and belongs to the technical field of image semantic segmentation. BACKGROUND

[0002] Remote sensing high-resolution image segmentation is an important part of remote sensing image processing, that is, automatically predicting each pixel of each scene image, and is a pixel-level classification task. Remote sensing high-resolution image segmentation is widely applied in the fields of city planning, land use, disaster assessment and environment monitoring. Early traditional remote sensing image segmentation mainly uses hand-designed classifiers based on edge, texture, shape, color and other bottom features to perform segmentation, and then performs semantic labeling on the segmented image. For example, a threshold-based segmentation method, a region growing-based segmentation method, an edge-based segmentation method and the like. These traditional methods can achieve good segmentation results to a certain extent, but different hand-designed feature extractors need to be designed according to the characteristics of different scenes. Since the requirement for hand-designed feature extractors is relatively high and the generalization performance is not good, the hand-designed feature extractors cannot be applied to real complex scenes on a large scale.

[0003] With the development of computers, the research in the field of artificial intelligence such as deep learning has also been broken through. The improvement of the operation capacity and transmission speed of computer hardware such as GPU and memory provides great impetus for the development of deep learning. The proposal of AlexNet brings the convolutional neural network back to people's vision. Semantic segmentation is a typical task in computer vision. In recent years, various practical architectures of semantic segmentation architecture have appeared, such as the fully convolutional network of FCNs and the encoder-decoder structure of DeepLabV3. The most commonly used is the encoder-decoder structure. Specifically, the encoder uses a pre-trained backbone network such as VGG and ResNet to extract convolutional features and obtain semantic information; the decoder convolves and up-samples the obtained feature map to the size of the original image to obtain the final segmentation image. With the help of high-performance GPU, a network model with higher precision can be trained, and the image segmentation method based on deep learning has become the mainstream method of remote sensing image segmentation. However, high-performance GPU is often high in cost, and although the general model has high precision, the inference speed is slow and cannot meet the requirements of high efficiency and rapidness.

[0004] ERFNet is one of the models with high accuracy and fast speed in semantic segmentation at present, which uses the mainstream encoder-decoder structure. In the encoder, in order to ensure the inference speed, the standard 2D convolution is replaced by two 3x1 and 1x3 1D convolutions using the convolution decomposition principle; the 3x3 convolution with a step of 2 and the maximum pooling are used in parallel to down-sample to obtain more feature information; at the same time, the dilated convolution is used to obtain more context information without changing the size of the feature map. The decoder also uses the decomposed convolution, and the deconvolution with a step of 2 is used for up-sampling. At the same time, in order to prevent overfitting, the neuron random inactivation (Dropout) strategy is used for regularization in the network. However, due to the insufficient number of layers of the ERFNet network model, the shallow features and deep features are not connected, and the context information cannot be fully obtained, which limits the segmentation ability of the network to small objects in the image, and at the same time, the deconvolution cannot correctly restore the image information, which limits the overall performance of the network. SUMMARY

[0005] The purpose of the present application is to overcome the deficiencies in the prior art, provide a remote sensing image segmentation method and device based on an improved ERFNet network, and further improve the segmentation ability of the model to effectively segment high-resolution remote sensing images.

[0006] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0007] In a first aspect, the present application provides a remote sensing image segmentation method based on an improved ERFNet network, comprising:

[0008] obtaining a remote sensing image;

[0009] inputting the remote sensing image into the trained improved ERFNet semantic segmentation network model to obtain a remote sensing image segmentation result, wherein:

[0010] the improved ERFNet semantic segmentation network model comprises an encoder module with an attention pyramid pooling module and a decoder module with a light flow guided alignment module instead of deconvolution.

[0011] Further, the step of inputting the remote sensing image into the trained improved ERFNet semantic segmentation network model to obtain a remote sensing image segmentation result comprises:

[0012] after the remote sensing image is input into the trained improved ERFNet semantic segmentation network model, it first enters the encoder module;

[0013] the encoder module extracts shallow features and deep features from the remote sensing image through down-sampling, standard convolution and dilated convolution to obtain extracted shallow feature maps and deep feature maps;

[0014] The attention pyramid pooling module is used to further extract features and context information from the extracted deep feature map, and a further processed deep feature map is obtained.

[0015] The further processed deep feature map and the shallow feature map are input into the optical flow guided alignment module in the decoder module for up-sampling, then 1x1 convolution is used to process the channel, and bilinear interpolation up-sampling is performed to obtain the remote sensing image segmentation result.

[0016] Further, the training method of the improved ERFNet semantic segmentation network model comprises:

[0017] The remote sensing high-resolution image dataset is obtained and data preprocessing is performed, and the dataset is divided into a training set, a validation set and a test set;

[0018] The pre-constructed improved ERFNet semantic segmentation network model is trained and verified using the training set and the validation set;

[0019] The test set is input into the trained improved ERFNet semantic segmentation network model to obtain a remote sensing high-resolution image segmentation result.

[0020] Further, the remote sensing high-resolution image dataset is obtained and data preprocessing is performed, and the dataset is divided into a training set, a validation set and a test set, comprising:

[0021] The remote sensing high-resolution image dataset is downloaded from an open source dataset website or self-made;

[0022] The pictures and labels in the remote sensing high-resolution image dataset are scaled in size and data enhanced;

[0023] The scaled and data enhanced pictures and labels are placed in different folders, and the training set, the validation set and the test set are divided according to the proportion.

[0024] Further, the pre-constructed improved ERFNet semantic segmentation network model is trained and verified using the training set and the validation set, comprising:

[0025] S21: Set the initial parameters of the training improved ERFNet semantic segmentation network model;

[0026] S22: In the training process, the neuron random inactivation strategy is used for regularization, and the cross-entropy loss function is used to calculate the loss;

[0027] S23: Calculate the gradient according to the loss obtained in S22, use the Adam optimizer to update the neural network weights and bias, and use the poly strategy for learning rate decay;

[0028] S24: Use the mean intersection over union mIoU to evaluate the model;

[0029] S25: repeating the training process of steps S22-S24, evaluating the network model using the validation set after each round of training is completed, until the training rounds are completed, and saving the optimal trained model.

[0030] In a second aspect, the present application provides a remote sensing image segmentation device based on an improved ERFNet network, comprising:

[0031] An acquisition module is configured to acquire a remote sensing image.

[0032] A segmentation module is configured to input the remote sensing image into the trained improved ERFNet semantic segmentation network model to obtain a remote sensing image segmentation result, wherein:

[0033] The improved ERFNet semantic segmentation network model comprises an encoder module with an attention pyramid pooling module added and a decoder module with a light flow guided alignment module replacing deconvolution.

[0034] Further, the segmentation module comprises:

[0035] An input unit is configured to input the remote sensing image into the trained improved ERFNet semantic segmentation network model first into the encoder module.

[0036] A first extraction unit is configured to extract shallow features and deep features of the remote sensing image by downsampling, standard convolution and dilated convolution in the encoder module to obtain extracted shallow feature maps and deep feature maps.

[0037] A second extraction unit is configured to further extract features and context information from the extracted deep feature maps using the attention pyramid pooling module to obtain further processed deep feature maps.

[0038] A segmentation unit is configured to input the further processed deep feature maps and shallow feature maps into the light flow guided alignment module for upsampling in the decoder module, then use 1x1 convolution to process channels and perform bilinear interpolation upsampling to obtain the remote sensing image segmentation result.

[0039] Further, the segmentation module is provided with a training module configured to train the improved ERFNet semantic segmentation network model, and the training module comprises:

[0040] A download unit is configured to download or self-make a remote sensing high-resolution image dataset from an open source dataset website.

[0041] A scaling unit is configured to perform size scaling and data enhancement on pictures and labels in the remote sensing high-resolution image dataset.

[0042] A dividing unit is configured to put the scaled and data-enhanced pictures and labels into different folders, and divide the training set, the verification set and the test set according to the proportions.

[0043] In a third aspect, the present application provides an electronic device, comprising a processor and a storage medium;

[0044] The storage medium is configured to store instructions;

[0045] The processor is configured to operate according to the instructions to perform the steps of the method according to any one of the preceding aspects.

[0046] In a fourth aspect, the present application provides a computer-readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the steps of the method according to any one of the preceding aspects.

[0047] Compared with the prior art, the present application has the following beneficial effects:

[0048] (1) The present application provides a remote sensing image segmentation method based on an improved ERFNet network, which adds an attention pyramid pooling module in the encoder and uses an optical flow guided alignment module instead of deconvolution in the decoder, fully fuses multi-scale information, ensures the image restoration accuracy, and effectively improves the segmentation accuracy of small objects and overall images in remote sensing high-resolution images.

[0049] (2) The remote sensing image segmentation method based on the improved ERFNet network used in the present application has the characteristics of high segmentation accuracy, small calculation parameter quantity and fast inference speed, and the remote sensing image segmentation method based on the improved ERFNet network can fuse multi-scale features, integrate global and local context information, guide image restoration by upsampling, and effectively improve the segmentation accuracy of small objects and overall images in remote sensing high-resolution images. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a training flowchart of the improved ERFNet semantic segmentation network model provided by the embodiments of the present application;

[0051] Figure 2 is a schematic diagram of the improved ERFNet semantic segmentation network provided by the embodiments of the present application;

[0052] Figure 3 is a schematic diagram of the down-sampling layer and the convolution block provided by the embodiments of the present application;

[0053] Figure 4 is a schematic diagram of the attention pyramid pooling module provided by the embodiments of the present application;

[0054] Figure 5 is a schematic diagram of the optical flow guided alignment module provided by the embodiments of the present application;

[0055] Figure 6 This is a flowchart of a remote sensing image segmentation method based on an improved ERFNet network provided in an embodiment of the present invention. Detailed Implementation

[0056] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0057] Example 1

[0058] like Figure 6 As shown, this embodiment introduces a remote sensing image segmentation method based on an improved ERFNet network, including:

[0059] Acquire remote sensing images;

[0060] The remote sensing image is input into the trained improved ERFNet semantic segmentation network model to obtain the remote sensing image segmentation result, where:

[0061] The improved ERFNet semantic segmentation network model includes an encoder module with added attention pyramid pooling and a decoder module that replaces deconvolution with optical flow-guided alignment.

[0062] The remote sensing image segmentation method based on the improved ERFNet network provided in this embodiment involves the following steps in its application process:

[0063] Step S1: Obtain the remote sensing high-resolution image dataset and perform data preprocessing to obtain the training set, validation set, and test set;

[0064] Step S2: Use the training set and validation set obtained in step S1 to train and validate the pre-built improved ERFNet semantic segmentation network model;

[0065] Step S3: Input the test set obtained in step S1 into the trained improved ERFNet semantic segmentation network model to obtain the remote sensing high-resolution image segmentation results.

[0066] Specifically, step S1 includes the following steps:

[0067] S11: Download or create your own remote sensing dataset from an open-source dataset website;

[0068] S12: Resize and augment the images and their labels in the downloaded dataset;

[0069] S13: Place the scaled and augmented images and labels into different folders, and divide them into training, validation and test sets according to the proportions.

[0070] Specifically, the improved ERFNet semantic segmentation network pre-constructed in step S2 includes an encoder adding an attention pyramid pooling module (APPM) and a decoder replacing deconvolution with a flow-guided alignment module (FGAM);

[0071] In the encoder module, shallow features and deep features of the high-resolution remote sensing image are extracted through downsampling, standard convolution and dilated convolution, to obtain the extracted shallow feature map and deep feature map; the attention pyramid pooling module is used to further extract features and context information from the extracted deep feature map, to obtain the further processed deep feature map;

[0072] In the decoder module, the further processed deep feature map and the shallow feature map are input into the flow-guided alignment module for upsampling, and then 1x1 convolution is used to process the channel and perform bilinear interpolation upsampling to obtain the final segmentation prediction map.

[0073] Specifically, step S2 includes the following steps:

[0074] S21: setting initial parameters for training the improved ERFNet semantic segmentation network model;

[0075] S22: using the neuron random inactivation (Dropout) strategy for regularization in the training process, and using the cross-entropy loss function to calculate the loss;

[0076] S23: calculating the gradient according to the loss obtained in S22, using the Adam optimizer to update the neural network weights and bias, and using the poly strategy for learning rate decay;

[0077] S24: using the mean intersection over union mIoU to evaluate the model;

[0078] S25: repeating the training process of steps S22-S24, evaluating the network model using the validation set after each training round, until the training rounds are completed, and saving the optimal training model.

[0079] Specifically, step S3 includes the following steps:

[0080] S31: importing the optimal model obtained in step S2, reading in the test set images and labels obtained in S1;

[0081] S32: calculating the mIoU score and saving the test results.

[0082] The contents designed in the above embodiments will be described below in conjunction with a preferred embodiment.

[0083] As Figure 1As shown, the embodiment provides a remote sensing image segmentation method based on an improved ERFNet network. The specific steps are as follows:

[0084] Step S1: Obtain a remote sensing high-resolution image dataset and perform data preprocessing to obtain a training set, a validation set and a test set;

[0085] Step S1 specifically includes the following steps:

[0086] S11: Download the Gaofen Image Dataset (GID) dataset from the open source dataset website http: / / captain.whu.edu.cn / WUDA-RSImg / gid.html;

[0087] S12: Scale 150 pictures and their labels in the downloaded dataset using the resize function in the torchvision.transforms toolkit, to a size of 1024*2048; perform data augmentation on the scaled dataset using horizontal flipping and translation methods;

[0088] S13: Place the scaled pictures and labels in different folders, and randomly divide the data in the dataset into a training set, a validation set and a test set according to a 3:1:1 ratio. The file name list files after division are placed in the path of the project, and are train.txt, val.txt and test.txt respectively.

[0089] Step S2: Train and validate the pre-constructed improved ERFNet semantic segmentation network model using the training set and the validation set obtained in step S1;

[0090] The pre-constructed improved ERFNet semantic segmentation network is as shown in Figure 2 It includes an encoder module and a decoder module. The encoder module adds an attention pyramid pooling module (APPM), and the decoder module uses a flow-guided alignment module (FGAM) instead of deconvolution for upsampling;

[0091] In the encoder module, the input remote sensing high-resolution image is downsampled to obtain a feature map F1 using the downsample layer shown in Figure 3 (left), and then the feature map F1 is downsampled using the downsample layer, and the feature map F2 is obtained using Figure 3The convolutional block shown on the right performs 5 feature extractions on the downsampled feature map (with a dilation rate of 1 for all iterations) to obtain feature map F2; the downsampling layer is used to downsample feature map F2, and the convolutional block performs 8 feature extractions on the downsampled feature map (with dilation rates of 2, 4, 8, 16, 2, 4, 8, 16 respectively) to obtain feature map F3; F1, F2, and F3 can be expressed by the formula:

[0092] F1 = Down(input)

[0093] F2 = Conv(Down(F1))

[0094] F3 = Conv(Down(F2))

[0095] Where Down represents the downsampling operation of the downsampling layer, Conv represents the convolution operation of the convolution block, and input represents the input high-resolution remote sensing image. The following is used: Figure 4 The attention pyramid pooling module shown performs global average pooling (GAP) on feature map F3, and then performs a 1×1 convolution on the GAP-pooled feature map to obtain feature map F4 containing global information. Feature map F3 is downsampled three times using average pooling, and then each downsampled feature map is extracted using 7×7, 5×5, and 3×3 convolutions to obtain feature maps F5, F6, and F7. Feature map F7 is upsampled using bilinear interpolation and added element-wise to feature map F6. The upsampled feature map is then added to feature map F5, and the upsampled feature map is then obtained to obtain attention feature map F8. Attention feature map F8 is multiplied element-wise with feature map F3, and then added to feature map F4 to obtain output feature map F9. F4, F5, F6, F7, F8, and F9 can be expressed using the following formula:

[0096] F4 = Conv 1×1 (GAP(F3))

[0097] F5 = Conv 7×7 (Avgpool(F3))

[0098] F6 = Conv 5×5 (Avgpool(Avgpool(F3)))

[0099] F7 = Conv 3×3 (Avgpool(Avgpool(Avgpool(F3))))

[0100]

[0101]

[0102] Among them, Conv1×1 Conv 3×3 Conv 5×5 Conv 7×7 are 1x1, 3x3, 5x5 and 7x7 convolution respectively, GAP is global average pooling, Avgpool is average pooling, Up is bilinear interpolation up-sampling, is element-wise addition, is element-wise multiplication.

[0103] In the decoder module, the deep feature map F9 and the shallow feature map F2 are input to the optical flow guided alignment module as shown in Figure 5 . The optical flow guided alignment module generates the optical flow field Flow1 by concatenating the up-sampled deep feature map F9 with the shallow feature map F2 and 3x3 convolution. The optical flow field Flow1 is used to guide the up-sampling of the feature map F9, and the channel attention Ch ann_Atten (including a global average pooling and two 1x1 convolutions) is used to further enhance the features of the up-sampled feature map F10. The feature map F10 is obtained; then the feature map F10 and the feature map F1 obtained by the encoder are repeated the above process to obtain the feature map F11. The feature map F11 is converted to the class number by using 1x1 convolution, and then up-sampled by bilinear interpolation to obtain the final segmentation prediction map Output. Flow1, F10, Flow2, F11, Ch ann_Atten and Output can be represented by the following formulas:

[0104] Flow1=Conv 3×3 (Cat(F2, Up(F9)))

[0105] F10=Flow_Up(Flow1, F9)

[0106] Flow2=Conv 3×3 (Cat(F1, Up(F10)))

[0107] F11=Flow_Up(Flow2, F9)

[0108] Ch ann_Atten=Conv 1×1 (Conv 1×1 (GAP(F)))

[0109] Output=Up(Conv 1×1 (F11))

[0110] Wherein, F can be F10 or F11, Cat is feature concatenation, and Flow_Up is optical flow guided up-sampling.

[0111] Step S2 specifically comprises the following steps:

[0112] S21: Set the initial parameters of the training improved ERFNet semantic segmentation network model as follows;

[0113] Initial learning rate: 0.0005;

[0114] Optimizer: Adam optimizer;

[0115] Learning rate decay: poly strategy;

[0116] Weight decay: 0.0002;

[0117] Momentum: 0.9;

[0118] Batch size: 6;

[0119] Training epochs: 300;

[0120] S22: Use the neuron random inactivation (Dropout) strategy for regularization during training, and use the cross-entropy loss function to calculate the loss, the calculation formula of the loss is as follows:

[0121] L = -[ylogy' + (1-y)log(1-y')]

[0122] Where y is the sample label, and y' is the predicted value. When y = 1, L = -logy', at this time, the closer y' is to 1, the smaller the loss value, and the closer y' is to 0, the larger the loss value. When y = 0, L = -log(1-y'), at this time, the closer y' is to 1, the larger the loss value, and the closer y' is to 0, the smaller the loss value.

[0123] S23: Calculate the gradient according to the loss obtained in S22, and update the neural network weights and biases using the Adam optimizer;

[0124] S24: Use the poly strategy for learning rate decay, and the decay formula is as follows:

[0125]

[0126] Where base_lr is the initial learning rate, cur_epoch is the current training epoch, and epoch is the total number of epochs.

[0127] S25: Evaluate the model using the average intersection over union (mIoU), which represents the accuracy of the network model segmentation. The higher the mIoU value, the better the image segmentation effect. The calculation method is as follows:

[0128]

[0129] where N represents the number of classes, i represents the ith class; TP i is the true positive, that is, the model's ith class prediction and the true class are both positive cases; FP i is the false positive, that is, the model's ith class prediction is positive, but the true class is negative; FN i is the false negative, that is, the model's ith class prediction is negative, but the true class is positive;

[0130] S26: Repeat the training process of steps S22-S25, evaluate the network model using the validation set after each training round, save the network parameters according to the optimal mIoU value, and save the optimal training model until the training round number reaches 300.

[0131] Step S3: Input the test set obtained in step S1 into the improved ERFNet semantic segmentation network obtained in step S2 to obtain the remote sensing high-resolution image segmentation result.

[0132] Step S3 specifically includes the following steps:

[0133] S31: Import the optimal model obtained in step S2, read in the pictures and labels of the test set obtained in S1;

[0134] S32: Calculate the mIoU score and save the test results.

[0135] Example 2

[0136] The embodiment provides a remote sensing image segmentation device based on an improved ERFNet network, which comprises:

[0137] An acquisition module is configured to acquire a remote sensing image.

[0138] A segmentation module is configured to input the remote sensing image into a trained improved ERFNet semantic segmentation network model to obtain a remote sensing image segmentation result, wherein:

[0139] The improved ERFNet semantic segmentation network model comprises an encoder module with an attention pyramid pooling module and a decoder module with a light flow guided alignment module instead of deconvolution.

[0140] In further embodiments, the segmentation module comprises:

[0141] The input unit is configured to input a remote sensing image into the improved ERFNet semantic segmentation network model, and then input the remote sensing image into an encoder module;

[0142] The first extraction unit is configured to extract shallow features and deep features of the remote sensing image by downsampling, standard convolution and dilated convolution in the encoder module, to obtain extracted shallow feature maps and deep feature maps;

[0143] The second extraction unit is configured to further extract features and context information from the extracted deep feature maps by using an attention pyramid pooling module, to obtain further processed deep feature maps;

[0144] The segmentation unit is configured to input the further processed deep feature maps and the shallow feature maps into an optical flow guided alignment module for upsampling in a decoder module, and then use 1x1 convolution to process channels and perform bilinear interpolation upsampling, to obtain a remote sensing image segmentation result.

[0145] In a further embodiment, the segmentation module is provided with a training module configured to train the improved ERFNet semantic segmentation network model, and the training module comprises:

[0146] The downloading unit is configured to download or self-make a remote sensing high-resolution image dataset from an open source dataset website;

[0147] The scaling unit is configured to perform size scaling and data enhancement on pictures and labels in the remote sensing high-resolution image dataset;

[0148] The division unit is configured to put the pictures and labels after scaling and data enhancement into different folders, and divide a training set, a validation set and a test set according to a proportion.

[0149] Embodiment 3

[0150] The embodiment provides an electronic device, comprising a processor and a storage medium;

[0151] The storage medium is configured to store instructions;

[0152] The processor is configured to operate according to the instructions to perform the steps of the method according to any one of the preceding embodiments.

[0153] Embodiment 4

[0154] The embodiment provides a computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the steps of the method according to any one of the preceding embodiments.

[0155] The above merely describes the preferred embodiments of the present application, and it should be pointed out that, for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present application, and these improvements and modifications should also be considered as the protection scope of the present application.

Claims

1. A remote sensing image segmentation method based on an improved ERFNet network, characterized in that, include: Acquire remote sensing images; The remote sensing image is input into the trained improved ERFNet semantic segmentation network model to obtain the remote sensing image segmentation result, where: The improved ERFNet semantic segmentation network model includes an encoder module with added attention pyramid pooling and a decoder module that replaces deconvolution with optical flow-guided alignment. The process of inputting remote sensing images into a trained, improved ERFNet semantic segmentation network model to obtain remote sensing image segmentation results includes: After the remote sensing image is input into the trained improved ERFNet semantic segmentation network model, it first enters the encoder module; In the encoder module, shallow and deep features are extracted from the remote sensing image through downsampling, standard convolution, and dilated convolution, resulting in the extracted shallow and deep feature maps. An attention pyramid pooling module is used to further extract features and contextual information from the extracted deep feature map, resulting in a further processed deep feature map. The decoder module inputs the further processed deep feature map and shallow feature map into the optical flow-guided alignment module for upsampling, then uses 1×1 convolution to process the channel and performs bilinear interpolation upsampling to obtain the remote sensing image segmentation result.

2. The remote sensing image segmentation method based on the improved ERFNet network according to claim 1, characterized in that, The training method for the improved ERFNet semantic segmentation network model includes: Acquire a high-resolution remote sensing image dataset and perform data preprocessing, dividing it into training set, validation set and test set; The pre-built improved ERFNet semantic segmentation network model was trained and validated using the training set and validation set; The test set is input into the trained improved ERFNet semantic segmentation network model to obtain the segmentation results of remote sensing high-resolution images.

3. The remote sensing image segmentation method based on the improved ERFNet network according to claim 2, characterized in that, The process of acquiring a high-resolution remote sensing image dataset and performing data preprocessing, dividing it into a training set, a validation set, and a test set, includes: Download or create your own high-resolution remote sensing image datasets from open-source dataset websites; Resizing and data augmentation are performed on images and labels in a remote sensing high-resolution image dataset; The scaled and augmented images and labels were placed in different folders and divided into training, validation and test sets according to proportions.

4. The remote sensing image segmentation method based on the improved ERFNet network according to claim 2, characterized in that, The process of training and validating a pre-built improved ERFNet semantic segmentation network model using a training set and a validation set includes: S21: Set the initial parameters for training the improved ERFNet semantic segmentation network model; S22: During training, a random neuron deactivation strategy is used for regularization, and the loss is calculated using the cross-entropy loss function; S23: Calculate the gradient based on the loss obtained in S22, update the neural network weights and biases using the Adam optimizer, and use the poly strategy to decay the learning rate. S24: Evaluate the model using the mean intersection-union ratio mIoU; S25: Repeat steps S22-S24 for training. After each training round, use the validation set to evaluate the network model until all training rounds are completed, and save the optimal training model.

5. A remote sensing image segmentation device based on an improved ERFNet network, characterized in that, include: The acquisition module is used to acquire remote sensing images; The segmentation module is used to input remote sensing images into the trained improved ERFNet semantic segmentation network model to obtain remote sensing image segmentation results, wherein: The improved ERFNet semantic segmentation network model includes an encoder module with added attention pyramid pooling and a decoder module that replaces deconvolution with optical flow-guided alignment. The segmentation module includes: The input unit is used for: after the remote sensing image is input into the trained improved ERFNet semantic segmentation network model, it first enters the encoder module; The first extraction unit is used to: extract shallow and deep features from remote sensing images in the encoder module through downsampling, standard convolution, and dilated convolution, and obtain the extracted shallow and deep feature maps. The second extraction unit is used to: further extract features and contextual information from the extracted deep feature map using the attention pyramid pooling module to obtain a further processed deep feature map; The segmentation unit is used to: input the further processed deep feature map and shallow feature map into the optical flow-guided alignment module for upsampling in the decoder module, then process the channel using 1×1 convolution, and perform bilinear interpolation upsampling to obtain the remote sensing image segmentation result.

6. The remote sensing image segmentation device based on the improved ERFNet network according to claim 5, characterized in that, The segmentation module includes a training module for training an improved ERFNet semantic segmentation network model. The training module includes: The download unit is used to download high-resolution remote sensing image datasets from open-source dataset websites or to create your own. The scaling unit is used to scale and augment images and labels in a remote sensing high-resolution image dataset. The partitioning unit is used to place the scaled and augmented images and labels into different folders and divide them into training, validation and test sets according to the proportions.

7. An electronic device, characterized in that: Including processor and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1 to 4.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 4.

Images