An intelligent sea ice monitoring method for remote sensing images based on wavelet transform and separable convolution semantic segmentation

Abstract

The present invention discloses a remote sensing image sea ice intelligent monitoring method based on wavelet transform and separable convolution semantic segmentation. Compared with other image segmentation methods, the present invention effectively improves the segmentation accuracy of sea ice remote sensing images. In the present invention, a deep encoding and decoding network model is built, and the use of deeper encoding layers can extract deeper semantic information, which helps to improve the segmentation accuracy. At the same time, the use of deep separable convolution to implement the encoding and decoding block can effectively reduce the increase in computational complexity caused by deepening the network. The shallow detail information of the image is extracted by wavelet transform, and the feature extracted by the shallow coding is merged to assist the network in feature extraction. The present invention can be used for tasks that require sea ice monitoring, such as marine aquaculture, environmental protection, sea surface operations and sea ice disaster prevention.

Description

Technical Field

[0001] The invention belongs to the field of sea ice remote sensing monitoring method research and semantic segmentation, and in particular to a remote sensing image sea ice intelligent monitoring method based on wavelet transform and separable convolution semantic segmentation. Technical Background

[0002] Sea ice is mainly distributed in polar waters and high-latitude waters. During its formation, sea ice will drive vertical convection of seawater, transporting seawater with high dissolved oxygen in the surface layer downward, and transporting fertile seawater rich in nutrients needed by phytoplankton to the surface, which is conducive to the mass reproduction of organisms and the formation of rich fishery resources. Sea ice can also reduce the tidal range and flow rate of tides and reduce the height of waves. When sea ice exists, the heat exchange between seawater and the atmosphere is greatly reduced. At the same time, sea ice has low thermal conductivity and high solar radiation reflectivity, which is conducive to restricting the change of ocean temperature. In addition, sea ice moves under the influence of sea breeze and seawater flow, which will cause sea ice disasters, affect the implementation of human activities on the coast and at sea and the safe operation of facilities, threaten life safety, and cause resource and property losses. Therefore, monitoring the distribution of sea ice is of great value to the implementation of marine aquaculture, environmental protection, and sea operations.

[0003] At present, the monitoring methods of sea ice are mainly divided into visual observation, instrument measurement and remote measurement. Among them, visual observation and instrument measurement are limited in capability and cannot realize the observation of the entire sea ice area, while remote measurement can realize large-scale sea ice monitoring through remote sensing satellites. However, due to the limitations of satellite hardware equipment and shooting environment, remote sensing images cannot show the details of sea surface objects, and it is difficult to accurately distinguish the sea ice area. Moreover, remote sensing satellites are easily affected by bad weather, clouds and fog during the shooting process. The images taken will have multiple interference targets and uneven distribution of sea ice areas and non-sea ice areas, which greatly increases the difficulty of identifying sea ice areas. In summary, the use of technical means to accurately monitor the sea ice area in sea ice remote sensing images is of great practical significance. Summary of the invention

[0004] The purpose of the present invention is to provide a method for intelligent monitoring of sea ice in remote sensing images based on wavelet transform and separable convolution semantic segmentation. Specifically, the sea ice area in the remote sensing image is segmented using a semantic segmentation network of a codec structure. By using deep separable convolution in the codec structure, the increase in computational complexity brought about by deepening the network can be balanced while keeping the convolution effect unchanged; for difficult-to-distinguish targets in sea ice remote sensing images, wavelet transform is used to extract detailed information in the horizontal, vertical and diagonal directions of the image, and the network is assisted in feature extraction. The channel attention mechanism is used to correct the sea ice features extracted in the encoding process, eliminate worthless features, and improve the accuracy of segmentation. Through verification on the remote sensing data set of sea ice target monitoring using the Ocean No. 1 satellite optical data, the present invention can effectively monitor the sea ice area.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] The remote sensing image sea ice intelligent monitoring method based on wavelet transform and separable convolution semantic segmentation specifically includes the following steps:

[0007] Step 1: Preprocess the original dataset of sea ice remote sensing images. The images in the dataset are in tif format and contain three channels: RGB. The images are divided into 256×256 images to obtain an expanded dataset, which is divided into a training set and a validation set at a ratio of 4:1. The test set will be given separately without processing.

[0008] Step 2: Build a semantic segmentation network model. Specifically, the built network model consists of four stages: wavelet transform stage, encoding stage, channel attention stage and decoding stage. The specific operations of each stage are as follows:

[0009] (1) Wavelet transform stage: A two-dimensional wavelet transform is performed on each channel of the input image using the Daubechies wavelet function with N=2. Each channel obtains a set of (4) transformed matrices, including: approximate matrix (LL), horizontal detail matrix (HL), vertical detail matrix (LH) and diagonal detail matrix (HH), whose sizes are 1 / 4 of the image before wavelet transform. The horizontal, vertical and diagonal detail matrices of the three RGB channels are stacked to obtain a merged detail matrix of size 128×128×9. Then the approximate matrix is ​​subjected to the same wavelet transform as above, and the generated detail matrices are stacked to obtain a merged detail matrix of size 64×64×9. Finally, the convolution (Conv) operation is used to increase the number of channels of the merged detail matrix to 64 and 128 respectively.

[0010] (2) Encoding stage: The encoding stage contains 5 serially connected encoder blocks (EncoderBlock). Depthwise separable convolution is used for feature extraction in each encoder block. The last layer in the encoder block is the pooling layer (MaxPooling2D). After each encoder block, the length and width of the feature map are halved and the number of channels is doubled. The input channel of the first encoder block is 32. Before pooling, the second and third encoder blocks respectively concatenate the feature maps with the corresponding size detail matrix output in the wavelet transform stage.

[0011] (3) Channel attention stage: Let the channel of the feature map input by the channel attention (ChannelAttention) be C. The channel attention performs global average pooling (GlobalAvggPool) and global maximum pooling (GlobalMaxPool) operations on the input feature map to obtain two 1×1×C compression matrices. The output compression matrices are input into two fully connected layers (Dense) in sequence. The first fully connected layer reduces the number of channels to C / 8, and the second fully connected layer restores the number of channels to C, obtaining two 1×1×C matrices. Then, the two matrices are added together and the hard-siggmoid activation function is used to obtain the channel weight matrix of the input feature map. Finally, the input feature map is dot-multiplied with the channel weight matrix and output.

[0012] (4) Decoding stage: The decoding stage contains 5 serially connected decoding blocks (DecoderBlock). Each decoding block is implemented using depthwise separable convolution, and the last layer in the decoding block is 2x upsampling (UpSampling2D). After each decoding block, the length and width of the feature map double, and the number of channels is halved. The input of each decoding block consists of two parts, namely the output of the previous decoding block and the output of the feature map of the encoding block at the same level after channel attention. The two are channel-wise spliced ​​as the input of the decoding block. At the end of the decoding stage, the softmax activation function is used to output the pixel-by-pixel prediction map.

[0013] Step 3: Set the parameters for model training. The number of training epochs is set to 150, the batch size is set to 16, the initial learning rate of the Adam optimizer is set to 0.001, and the "poly" learning rate strategy is adopted (power is set to 0.9), and the auxiliary loss weight is set to [0.16, 0.4, 1, 1].

[0014] Step 4: In view of the imbalance between the areas with sea ice and the areas without sea ice in the sea ice remote sensing dataset, the loss function adopts Focal Loss, and the specific formula is:

[0015] FL(p t )=-α t(1-p t ) γ log(p t );

[0016] Among them, p t Indicates the probability that the sample belongs to the true category, α t To control the shared weight of positive and negative samples on the total loss, when α t When the value is small, the weight of negative samples can be reduced; γ is the focusing parameter, which is a non-negative value, (1-p t ) γ It is called the modulation coefficient, which can reduce the weight of easy-to-classify samples, so that the model can focus more on difficult-to-classify samples during training. t Set it to 0.25 and gamma to 2.

[0017] Step 5: Input the processed data set into the built network model for training.

[0018] Step 6: During the test, read in a test image each time, take 256×256 blocks of the image for prediction, and splice all the prediction images back to the original image size. Use the spliced ​​prediction images and labels to calculate the weighted frequency intersection over union (FWIoU). The specific formula is:

[0019]

[0020] where p ii represents the number of pixels that are actually in the i-th category and predicted to be in the i-th category, p ij Indicates the number of pixels that are actually of the i-th class and predicted to be of the j-th class.

[0021] The remote sensing image sea ice intelligent monitoring method based on wavelet transform and separable convolution semantic segmentation of the present invention has the following beneficial effects:

[0022] (1) Provides an intelligent remote sensing image sea ice monitoring method;

[0023] (2) The network model uses a deep encoding and decoding structure, which can extract deeper semantic information. At the same time, the use of deep separable convolution to implement the encoding block and decoding block can effectively reduce the amount of calculation caused by deepening the network depth;

[0024] (3) The detailed information of the image after wavelet transformation is spliced ​​with the shallow feature map of the coding layer to supplement the shallow features extracted by the network, which helps to improve the accuracy of sea ice area segmentation;

[0025] (4) In order to address the imbalance between sea ice areas and non-sea ice areas in sea ice remote sensing images, the FocalLoss loss function is used to train the model to improve the recognition accuracy of sea ice areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of the network model structure of the present invention;

[0027] Figure 2 A schematic diagram of the network model wavelet transform accessing the encoder block (EncoderBlock) of the present invention;

[0028] Figure 3 It is a structural diagram of the network model encoder block (EncoderBlock) of the present invention;

[0029] Figure 4 It is a structural diagram of the network model channel attention (ChannelAttention) of the present invention;

[0030] Figure 5 It is a structural diagram of the network model decoding block (DecoderBlock) of the present invention. DETAILED DESCRIPTION

[0031] The invention will now be described in further detail with reference to the accompanying drawings.

[0032] Step 1: Preprocess the original dataset of sea ice remote sensing images. The images in the dataset are in tif format and contain three channels: RGB. The images are divided into 256×256 images to obtain an expanded dataset. The dataset is divided into a training set and a validation set in a ratio of 4:1. The test set will be given separately without processing.

[0033] Step 2: Build the network model of the present invention. Specifically, the built network model includes four stages, namely: wavelet transform stage, encoding stage, channel attention stage and decoding stage. The specific operations of each stage are as follows:

[0034] (1) Wavelet transform stage: A two-dimensional wavelet transform is performed on each channel of the input image using the Daubechies wavelet function with N=2. Each channel obtains a set of (4) transformed matrices, including: approximate matrix (LL), horizontal detail matrix (HL), vertical detail matrix (LH) and diagonal detail matrix (HH), each of which is 1 / 4 of the size of the image before the transform. The horizontal, vertical and diagonal detail matrices of the three RGB channels are stacked to obtain a merged detail matrix of size 128×128×9. Then the approximate matrix is ​​subjected to the same wavelet transform as above, and the generated detail matrices are stacked to obtain a merged detail matrix of size 64×64×9. Finally, the convolution operation is used to increase the number of channels of the merged detail matrix to 64 and 128 respectively.

[0035] (2) Encoding stage: The encoding stage contains 5 serially connected encoder blocks (EncoderBlock). Each encoder block uses depthwise separable convolution for feature extraction, and the last layer in the encoder block is the pooling layer (MaxPooling2D). After each encoder block, the length and width of the feature map are halved and the number of channels is doubled. The input channel of the first encoder block is 32. Before pooling, the second and third encoder blocks respectively concatenate the feature maps with the corresponding size detail matrix output in the wavelet transform stage. The depthwise separable convolution of the encoder block in this stage includes two branches. Let the number of input channels of the encoder block be C, which includes the following steps:

[0036] a) The first branch: This branch contains four layers of convolution. The convolution kernel of the first layer is 1×1×C, the convolution kernel of the second layer is 3×3×C, the convolution kernel of the third layer is 1×1×2C, and the convolution kernel of the fourth layer is 3×3×2C. The activation function is Relu, and 0 padding is used. This branch is finally processed by batch normalization (BN).

[0037] b) The second branch: This branch contains a layer of convolution, the convolution kernel is 1×1×2C, the activation function is Relu, and 0 padding is used. This branch is finally processed by batch normalization (BN).

[0038] c) Branch merging: The outputs of the two branches are added (ADD), passed through a ReLU activation function, and finally a maximum pooling layer with a 2×2 pooling kernel.

[0039] (3) Channel attention stage: Let the channel of the feature map input by channel attention (ChannelAttention) be C. Channel attention performs global average pooling (GlobalAvggPool) and global maximum pooling (GlobalMaxPool) operations on the input feature map to obtain two 1×1×C compression matrices. The output compression matrices are input into two fully connected layers (Dense) in sequence. The first fully connected layer reduces the number of channels to C / 8, and the second fully connected layer restores the number of channels to C, obtaining two 1×1×C matrices. Then, the two matrices are added together and the hard-sigmoid activation function is used to obtain the channel weight matrix of the input feature map. Finally, the input feature map is dot-multiplied with the channel weight matrix and output.

[0040] (4) Decoding stage: The decoding stage contains 5 serially connected decoding blocks (DecoderBlock). Each decoding block is implemented using depthwise separable convolution, and the last layer in the decoding block is 2x upsampling (UpSampling2D). After each decoding block, the length and width of the feature map are doubled and the number of channels is halved. The input of each decoding block consists of two parts, namely the output of the previous decoding block and the output of the feature map of the encoding block at the same level after channel attention. The two are channel-wise spliced ​​as the input of the decoding block. At the end of the decoding stage, the softmax activation function is used to output a pixel-by-pixel prediction map. The depthwise separable convolution of the decoding block in this stage includes two branches. Let the number of input channels of the decoding block be C, including the following steps:

[0041] a) The first branch: This branch contains three layers of convolution. The convolution kernel of the first layer is 1×1×C, the convolution kernel of the second layer is 3×3×C, and the convolution kernel of the third layer is The activation function is Relu, and 0 padding is used. This branch is finally batch normalized (BN).

[0042] b) The second branch: This branch contains a layer of convolution, and the convolution kernel is The activation function is Relu, and 0 padding is used. This branch is finally batch normalized (BN).

[0043] c) Branch merging: The outputs of the two branches are added (ADD), passed through a ReLU activation function, and finally a 2×2 kernel upsampling.

[0044] Step 3: Set the parameters for model training. The number of training epochs is set to 150; the batch size is set to 16; the initial learning rate of the Adam optimizer is set to 0.001, and the "poly" learning rate strategy is adopted (power is set to 0.9), and the auxiliary loss weight is set to [0.16, 0.4, 1, 1].

[0045] Step 4: In view of the imbalance between the areas with sea ice and the areas without sea ice in the sea ice remote sensing dataset, the Focal Loss function is used in the training process. The specific formula is:

[0046] FL(p t )=-α t (1-p t ) γ log(p t );

[0047] Among them, p t Indicates the probability that the sample belongs to the true category; α t To control the shared weight of positive and negative samples on the total loss, when α t When the value is small, the weight of negative samples can be reduced; γ is the focusing parameter, which is a non-negative value, (1-p t ) γ It is called the modulation coefficient, which can reduce the weight of easy-to-classify samples, so that the model can focus more on difficult-to-classify samples during training. t Set it to 0.25 and gamma to 2.

[0048] Step 5: Input the processed data set into the built network model for training.

[0049] Step 6: During the test, read in a test image each time, take 256×256 blocks of the image for prediction, and splice all the prediction images back to the original image size. Use the spliced ​​prediction images and labels to calculate FWIoU (weighted frequency intersection over union). The specific formula is:

[0050]

[0051] where p ii represents the number of pixels that are actually in the i-th category and predicted to be in the i-th category, p ij Indicates the number of pixels that are actually of the i-th class and predicted to be of the j-th class.

[0052] The effect of the present invention can be illustrated by the following simulation experiment:

[0053] The simulation experiment uses the Keras deep learning framework to build the network model, GeForce GTX 1080Ti for GPU accelerated training, and the operating system is a 64-bit Linux operating system.

[0054] The data set for the simulation experiment was given by the 5th "China Science Star Cup" International High-resolution Remote Sensing Image Interpretation Competition. The data set is the optical data of the Ocean No. 1 satellite. The scene covers the area around the Bohai Sea in my country, with a resolution of 50m, including two categories: sea ice and background. The data set contains three sizes of pictures: 512×512, 1024×1024, and 2048×2048. The remote sensing image is in tif format, containing three RGB channels; the annotation file is a single-channel png image with pixel values ​​of 0 and 255, corresponding to the background and sea ice, respectively. There are 1,500 training sets and more than 1,000 test sets.

[0055] In order to verify the advantages and benefits of the present invention, the present invention is compared with three existing image segmentation methods, namely SDFCN, DeepLabv3+ and ResNet50. For details, please refer to the following documents:

[0056] G.Chen, Sensing, vol.11, no.5, pp.1633-1644, May2018, doi: 10.1109 / JSTARS.2018.2810320.

[0057] Chen LC, Yukun Zhu, Papandreou G and SchroffE.Encoder-Decoder withAtrous Separable Convolution for Semantic Image Segmentation[C] / / The European Conference on Computer Vision(ECCV), Berlin: Springer, 2018, pp.801-818.

[0058] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jiian Sun.Deep residuallearning for image recognition[C] / / .Internaltional Conference on ComputerVision and Pattern Recogintion.NJ: IEEE, 2016, PP.770-778.

[0059] The weighted frequency intersection over union (FWIoU) is used as the evaluation index to evaluate the performance of the model. The larger the value, the better the performance. Since the amount of original data is too small, all images in the training set are cropped into blocks of 256×256, and the training set is expanded to 8112 images.

[0060] It can be observed from Table 1 that the weighted frequency intersection and union (WFIU) of the present invention on the sea ice remote sensing data set is about 1% higher than that of the other three methods when the training model converges. On the test set, the weighted frequency intersection and union (WFIU) of the present invention is 4.2%-5.5% higher than that of the other three methods. The main reasons why the present invention achieves better segmentation results are: 1) The network model of the present invention uses a deep encoding and decoding structure, which can extract deeper semantic information. At the same time, the use of deep separable convolution to implement the encoding block and the decoding block can effectively reduce the amount of calculation caused by deepening the network depth. 2) The detailed information of the image after wavelet transformation is fused with the shallow feature map of the encoding layer to supplement the shallow features extracted by the network, which helps to improve the accuracy of sea ice area segmentation. 3) In view of the imbalance between sea ice areas and sea ice-free areas in sea ice remote sensing images, the Focal Loss loss function is used to train the model, thereby improving the segmentation accuracy of sea ice areas.

[0061] Table 1 Comparison of objective evaluation indicators of the proposed method with SDFCN, DeepLabv3+, and ResNet50 on sea ice remote sensing dataset

[0062]

[0063] Those skilled in the art will appreciate that the above simulation experiments are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and implementation examples. Those skilled in the art can make various other specific variations and combinations based on the technical inspirations disclosed in the present invention without departing from the essence of the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A sea ice remote sensing image segmentation method based on deep separable convolution and wavelet transform assisted feature extraction, characterized in that: The method specifically comprises the following steps: Step 1: Preprocess the original data set of sea ice remote sensing images; Step 2: Build a semantic segmentation network model; Step 3: Initialize the parameters for model training; Step 4: Use the Focal Loss function during training; Step 5: Input and segment the data set to start training; Step 6: Output the predicted results of sea ice segmentation and calculate its weighted frequency intersection and union ratio with the annotated ones; Among them, the semantic segmentation network model built in step 2 consists of four stages: wavelet transform stage, encoding stage, channel attention stage and decoding stage. The specific operations of each stage are as follows: (1) Wavelet transform stage: A two-dimensional wavelet transform is performed on each channel of the input image using the Daubechies wavelet function with N=2. Each channel obtains a set of (4) transformed matrices, including: approximate matrix (LL), horizontal detail matrix (HL), vertical detail matrix (LH) and diagonal detail matrix (HH), whose sizes are all 1 / 4 of the size of the image before the transform; the horizontal, vertical and diagonal detail matrices of the three RGB channels are stacked to obtain a merged detail matrix of size 128×128×9; then the approximate matrix is ​​subjected to the same wavelet transform as above, and the generated detail matrices are stacked to obtain a merged detail matrix of size 64×64×9; finally, a convolution (Conv) operation is used to increase the number of channels of the merged detail matrix to 64 and 128 respectively; (2) Encoding stage: The encoding stage includes five serially connected encoder blocks (EncoderBlock). Each encoder block uses depthwise separable convolution for feature extraction. The last layer in the encoder block is the pooling layer (MaxPooling2D). After each encoder block, the length and width of the feature map are halved and the number of channels is doubled. The input channel of the first encoder block is 32. Before pooling, the second and third encoder blocks respectively concatenate the feature maps with the corresponding size detail matrix output in the wavelet transform stage. The depthwise separable convolution of the encoder block in this stage includes two branches. Let the number of input channels of the encoder block be C. It includes the following steps: a) The first branch: This branch contains four layers of convolution; the convolution kernel of the first layer is 1×1×C, the convolution kernel of the second layer is 3×3×C, the convolution kernel of the third layer is 1×1×2C, and the convolution kernel of the fourth layer is 3×3×2C. The activation function is Relu, and 0 padding is used. This branch is finally processed by batch normalization (BN); b) The second branch: This branch contains a layer of convolution, the convolution kernel is 1×1×2C, the activation function is Relu, and 0 padding is used. This branch is finally processed by batch normalization (BN); c) Branch merging: add the outputs of the two branches (ADD), pass through a Relu activation function, and finally a maximum pooling layer with a 2×2 pooling kernel; (3) Channel attention stage: Let the channel of the feature map input by channel attention (ChannelAttention) be C; channel attention performs global average pooling (GlobalAvgPool) and global maximum pooling (GlobalMaxPool) operations on the input feature map to obtain two 1×1×C compression matrices; the output compression matrices are input into two fully connected layers (Dense) in sequence, where the first fully connected layer reduces the number of channels to C / 8, and the second fully connected layer restores the number of channels to C, obtaining two 1×1×C matrices; then the two matrices are added and the hard-sigmoid activation function is used to obtain the channel weight matrix of the input feature map; finally, the input feature map is dot-multiplied with the channel weight matrix and output; (4) Decoding stage: The decoding stage includes five serially connected decoding blocks (DecoderBlock). Each decoding block is implemented using depthwise separable convolution, and the last layer in the decoding block is 2x upsampling (UpSampling2D). After each decoding block, the length and width of the feature map double and the number of channels is halved. The input of each decoding block consists of two parts, namely the output of the previous decoding block and the output of the feature map of the encoding block at the same level after channel attention. The two are channel-wise concatenated as the input of the decoding block. At the end of the decoding stage, the softmax activation function is used to output a pixel-by-pixel prediction map. In this stage, the depthwise separable convolution of the decoding block includes two branches. Let the number of input channels of the decoding block be C, which includes the following steps: a) The first branch: This branch contains three layers of convolution; the convolution kernel of the first layer is 1×1×C, the convolution kernel of the second layer is 3×3×C, and the convolution kernel of the third layer is The activation function is Relu, 0 padding is used, and batch normalization (BN) is performed at the end; b) The second branch: This branch contains a layer of convolution, and the convolution kernel is The activation function is Relu, 0 padding is used, and batch normalization (BN) is performed at the end; c) Branch merging: The outputs of the two branches are added (ADD), passed through a ReLU activation function, and finally a 2×2 kernel upsampling.

2. The sea ice remote sensing image segmentation method based on deep separable convolution and wavelet transform assisted feature extraction according to claim 1 is characterized in that: The images in the dataset are in tif format and contain three RGB channels. The images are divided into 256×256 images to obtain an expanded dataset. The dataset is divided into a training set and a validation set in a ratio of 4:

1. The test set will be given separately without processing.

3. The sea ice remote sensing image segmentation method based on deep separable convolution and wavelet transform assisted feature extraction according to claim 1 is characterized in that: In the model training parameters, the number of training epochs is set to 150; the batch size is set to 16; the initial learning rate of the Adam optimizer is set to 0.001, and the "poly" learning rate strategy is adopted (power is set to 0.9), and the auxiliary loss weight is set to [0.16, 0.4, 1, 1].

4. The sea ice remote sensing image segmentation method based on deep separable convolution and wavelet transform assisted feature extraction according to claim 1 is characterized in that: The Focal Loss loss function used in the training process has the following formula: FL(p t )=-a t (1-p t ) γ log(p t ); Among them, p t Indicates the probability that the sample belongs to the true category; α t To control the shared weight of positive and negative samples on the total loss, when α t When the value is small, the weight of negative samples can be reduced; γ is the focusing parameter, which is a non-negative value, (1-p t ) γ It is called the modulation coefficient, which can reduce the weight of easy-to-classify samples, so that the model can focus more on difficult-to-distinguish samples during training. t Set it to 0.25 and gamma to 2.

5. The sea ice remote sensing image segmentation method based on deep separable convolution and wavelet transform assisted feature extraction according to claim 1 is characterized in that: During prediction, a test image is read in each time, and 256×256 blocks of the image are taken in turn for prediction. All the predicted images are spliced ​​back to the original image size, and the FWIoU (weighted frequency intersection over union) is calculated using the spliced ​​predicted image and the label. The specific formula is: where p ii represents the number of pixels that are actually in the i-th category and predicted to be in the i-th category, p ij Indicates the number of pixels that are actually of the i-th class and predicted to be of the j-th class.

Images