A cloud image recognition method based on multi-scale features and gradient fusion operation
By fusing multi-scale features and gradient information in the cloud map recognition method, using the Canny algorithm to extract gradient information and perform feature fusion, the cloud map recognition problem in the existing technology is solved, and the cloud map classification performance and contour feature extraction capability are improved.
Patent Information
- Application Number
- CN202210192286.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-02-28
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2042-02-28
AI Technical Summary
The existing cloud map recognition methods fail to effectively integrate the gradient features and multi-scale features of cloud maps, resulting in difficulty in classifying cloud categories, especially the similar shapes of cloud maps in different categories and complex colors and shape characteristics of similar cloud maps.
The cloud map recognition method based on multi-scale features and gradient fusion operations is adopted. By reducing the size of cloud map features by half and adding multi-scale feature information, the gradient feature fusion module is combined with the gradient feature fusion module to extract gradient information, and transform it into a weight matrix and feature map for point multiplication. Operation to enhance the contour feature extraction capability.
It improves the cloud map classification performance, enhances the network's information extraction ability of cloud map contour features, and expands the application and research in the field of cloud map recognition.
Smart Images

Figure CN114565534B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of meteorological observation, and in particular to a cloud image recognition method based on multi-scale features and gradient fusion operations. Background Art
[0002] In recent years, cloud classification research has been widely used in meteorological monitoring, field observation, aerospace and other fields, so how to accurately classify cloud types has become a challenging task. At present, the most mainstream ground-based cloud observation equipment is the all-sky imager. For example, Taihu and Wuxi stations use this instrument to collect ground observation data to analyze the imaging characteristics of images under different weather conditions. Another commonly used device is a meteorological satellite. However, meteorological satellites simply provide spectral information of cloud images, which is not as intuitive as the images of all-sky imagers. Cloud image recognition belongs to fine-grained recognition, that is, it is necessary to specifically identify the category of each cloud. This classification task is extremely challenging because cloud images of different categories have relatively similar shapes, and clouds of the same category have complex and diverse color and shape characteristics.
[0003] There are current cloud recognition methods that use meteorological data information to assist cloud recognition, but there is no method that fuses and enhances the gradient features of cloud images with multi-scale cloud features to assist cloud recognition. Summary of the invention
[0004] In order to solve the above problems, the present invention provides a cloud image recognition method based on multi-scale features and gradient fusion operation.
[0005] The present invention is achieved through the following technical solutions:
[0006] A cloud image recognition method based on multi-scale features and gradient fusion operation, comprising:
[0007] 1) The feature size of the cloud image is halved by branching, and the multi-scale feature information of the cloud image is increased, including the following steps:
[0008] After the cloud map passes through a 7×7 convolution layer and a maximum pooling layer, the feature map size is halved through a 3×3 convolution, and then sent to the branch network for training. Finally, the features of the main network and the branch network are fused at the corresponding positions of the feature map size.
[0009] 2) Enhance the contour information of cloud image features through the gradient feature fusion module, including the following steps:
[0010] The Canny edge contour extraction algorithm is used to extract gradient information from the feature map after each dense block, and the gradient information is converted into a weight matrix through softmax. Finally, a dot multiplication operation is performed with the feature map to enhance the original feature map's ability to extract contour information.
[0011] The method specifically comprises the following steps:
[0012] S1: Obtain the type image data of the cloud image, divide the cloud image into N categories according to the type characteristics of the cloud image, convert the cloud image size into 294×294, and normalize the image;
[0013] S2: The cloud image is input into the network to extract image features. After convolution and maximum pooling layers, a 64-dimensional feature map of size 74×74 is obtained. The network is divided into a backbone network and a branch network. The backbone network passes through dense block (1-1), transition layer (1-1), dense block (1-2), transition layer (1-2), dense block (1-3), transition layer (1-3) and dense block (1-4) respectively. The branch network passes through dense block (2-1), transition layer (2-1), dense block (2-2), transition layer (2-2), dense block (2-3), transition layer (2-3) and dense block (2-4) respectively.
[0014] S3: The input of the backbone network, transition layer (1-1) and transition layer (1-2) are halved by 3×3 convolutions with a step size of 2, and then superimposed with the output of dense block (2-1), dense block (2-2) and dense block (2-3) of the branch network. The features after the maximum pooling layer are halved by multiple 3×3 convolutions, and then channel-superimposed with the features after the dense blocks of the branch network to retain the most original feature information.
[0015] S4: The Canny algorithm is used to extract the gradient information of the original image, and the gradient enhancement operation is performed with the features after the first dense block of the main line and branch network; the feature map gradient is used to enhance the contour features after the fourth dense block of the main line and branch network.
[0016] Furthermore, the feature map size of the backbone network is 1 times larger than that of the branch network at the same position.
[0017] Furthermore, S1 specifically includes the following steps:
[0018] S1.1: According to the different characteristics of cloud types, cloud images are divided into N categories;
[0019] S1.2: The image data pixels are resized to 294×294, and the cloud features are normalized to [-1, 1].
[0020] Furthermore, S2 specifically includes the following steps:
[0021] S2.1: Input normalized cloud feature map;
[0022] S2.2: The input image first passes through a convolution layer with a kernel of 7×7 and a stride of 2 to obtain a feature map of size 147×147. Batch normalization and ReLU activation are performed after each convolution layer.
[0023] S2.3: The obtained features are passed through the maximum pooling layer to obtain a feature map of size 74×74;
[0024] S2.4, the obtained features are input into the main network and the branch network respectively; the main network is input into the dense block (1-1) to obtain features of size 74×74; the branch network is subjected to a 3×3 convolution operation with a step size of 2 to obtain features of size 37×37, and then input into the dense block (2-1) to obtain features of size 37×37; the features obtained after each group of convolutions will be superimposed with the previously input features by the number of channels as the input for the next convolution;
[0025] S2.5: The features obtained by dense block (1-1) and dense block (2-1) are respectively input into transition layer (1-1) and transition layer (2-1) to obtain features of size 37×37 and 19×19;
[0026] S2.6: The obtained 37×37 features are then passed through dense block (1-2), transition layer (1-2), dense block (1-3), transition layer (1-3), and dense block (1-4); the obtained 19×19 features are then passed through dense block (2-2), transition layer (2-2), dense block (2-3), transition layer (2-3), and dense block (2-4); among which, the number of convolution groups of dense block (1-2), dense block (1-3), and dense block (1-4) are 12, 24, and 16 respectively, and the number of convolution groups of dense block (2-2), dense block (2-3), and dense block (2-4) are 12, 24, and 16 respectively; after the above operations, the main features of size 10×10 and the branch features of size 5×5 are obtained respectively;
[0027] S2.7: The features obtained by the backbone network are subjected to a 3×3 convolution with a stride of 2 to output 5×5 features;
[0028] S2.8, after unifying the feature map sizes of the backbone network and the branch network to 5×5, the number of feature channels is superimposed to obtain 2048 channels, and the features of size 1×1 are obtained through the global average pooling layer;
[0029] S2.9, the obtained features are passed through the classification layer to calculate the classification results, and the total loss of the network is calculated; among them, the classification layer includes a classification convolution layer and a softmax function, and the classification layer is a convolution with a step size of 1 and 10 categories.
[0030] Furthermore, the maximum pooling layer described in S2.3 uses a convolution kernel size of 3×3 and a step size of 2.
[0031] Furthermore, dense block (1-1) and dense block (2-1) in S2.4 each contain 6 groups of identical convolutions, each group of convolutions contains 1 convolution layer with a convolution kernel of 1×1 and 1 convolution layer with a convolution kernel of 3×3, and the step size is 1.
[0032] Furthermore, the transition layer in S2.5 includes a convolution layer with a convolution kernel of 1×1 and a stride of 1, and an average pooling layer with a 2×2 kernel and a stride of 2.
[0033] Furthermore, S3 specifically includes the following steps:
[0034] S3.1: The 74×74 size feature obtained in S2.3 is used as input. The input feature undergoes three 3×3 convolutions with a stride of 2 to generate features of sizes 37×37, 19×19, and 10×10 respectively.
[0035] S3.2: Superimpose the features of size 19×19 with the number of output channels of dense block (2-2), and superimpose the features of size 10×10 with the number of output channels of dense block (2-3);
[0036] S3.3: The output size of the maximum pooling layer is 74×74, which becomes 37×37 after a 3×3 convolution with a stride of 2, and is superimposed with the output channel number of the dense block (2-1). The output of the transition layer (1-1) undergoes a 3×3 convolution with a stride of 2 to obtain a feature size of 19×19, which is superimposed with the output channel number of the dense block (2-2). The output of the transition layer (1-2) undergoes a 3×3 convolution with a stride of 2 to obtain a feature size of 10×10, which is superimposed with the output channel number of the dense block (2-3).
[0037] Furthermore, S4 specifically includes the following steps:
[0038] S4.1: Using the cloud image input, the Canny edge contour detection operator is used to calculate the contour features of the original image input, and then the softmax calculation is added with 1 to convert it into a feature weight, which is multiplied by the output of dense block (1-1) and dense block (2-1) after the S3.3 operation;
[0039] S4.2: After calculating the contour features of the outputs of dense blocks (1-4) and dense blocks (2-4) using the Canny edge contour detection operator, the softmax calculation is added with 1 to convert them into feature weights, which are multiplied by the outputs of dense blocks (1-4) and dense blocks (2-4) as the final output features.
[0040] The beneficial effects of the present invention are:
[0041] 1. The cloud map recognition method based on gradient information and multi-scale feature fusion provided by the present invention not only integrates the multi-scale feature information of the cloud map, but also extracts cloud map gradient information of different scales as weights in the middle of the model, enhances the contour features of the cloud map, and improves the classification performance of the network.
[0042] 2. The cloud image recognition method based on gradient information and multi-scale feature fusion provided by the present invention can further expand the application and research in the field of cloud image recognition. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.
[0044] Attached Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION
[0045] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0046] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.
[0047] A cloud image recognition method based on multi-scale features and gradient fusion operation, comprising:
[0048] 1) The feature size of the cloud image is halved by branching, and the multi-scale feature information of the cloud image is increased, including the following steps:
[0049] After the cloud map passes through a 7×7 convolution layer and a maximum pooling layer, the feature map size is halved through a 3×3 convolution, and then sent to the branch network for training. Finally, the features of the main network and the branch network are fused at the corresponding positions of the feature map size.
[0050] 2) Enhance the contour information of cloud image features through the gradient feature fusion module, including the following steps:
[0051] The Canny edge contour extraction algorithm is used to extract gradient information from the feature map after each dense block, and the gradient information is converted into a weight matrix through softmax. Finally, a dot multiplication operation is performed with the feature map to enhance the original feature map's ability to extract contour information.
[0052] Refer to the attached Figure 1 As shown in Table 1, the method specifically includes the following steps:
[0053] S1: Obtain the type image data of the cloud image, divide the cloud image into N categories according to the type characteristics of the cloud image, convert the cloud image size into 294×294, and normalize the image;
[0054] Among them, S1 specifically includes the following steps:
[0055] S1.1, according to the different characteristics of cloud types, cloud images are divided into N categories;
[0056] S1.2, the image data pixels are resized to 294×294, and then the cloud features are normalized to [-1, 1].
[0057] S2: The cloud image is used as the input network to extract image features. After convolution and maximum pooling layers, a 64-dimensional feature map of size 74×74 is obtained. The network is divided into a backbone network and a branch network. The feature map size of the backbone network is 1 times larger than that of the branch network at the same position. The backbone network passes through dense block (1-1), transition layer (1-1), dense block (1-2), transition layer (1-2), dense block (1-3), transition layer (1-3) and dense block (1-4), respectively. The branch network passes through dense block (2-1), transition layer (2-1), dense block (2-2), transition layer (2-2), dense block (2-3), transition layer (2-3) and dense block (2-4) respectively.
[0058] Among them, S2 specifically includes the following steps:
[0059] S2.1, input the normalized cloud feature map, the feature map size is 294×294;
[0060] S2.2, the input image first passes through a convolution layer with a convolution kernel of 7×7 and a stride of 2 to obtain a feature map of size 147×147. Batch normalization (BN) and ReLU activation are performed after each convolution layer in the network framework;
[0061] S2.3, the obtained features are passed through the maximum pooling layer to obtain a feature map of size 74×74. The maximum pooling layer uses a convolution kernel size of 3×3 and a step size of 2;
[0062] S2.4, the features obtained are subjected to two operations, one is that the backbone network is input into the dense block (1-1) to obtain features of size 74×74; the other is that the branch network is subjected to a 3×3 convolution operation with a step size of 2 to obtain features of size 37×37, which are input into the dense block (2-1) to obtain features of size 37×37; wherein, the dense block (1-1) and the dense block (2-1) each contain 6 groups of identical convolutions, each group of convolutions contains 1 convolution layer with a convolution kernel of 1×1 and 1 convolution layer with a convolution kernel of 3×3, and the step size is 1, so the feature size is not reduced. The dense block contains a dense stacking operation, that is, the features obtained after each group of convolutions will be stacked with the features input before by the number of channels as the input of the next convolution.
[0063] S2.5, the features obtained by dense block (1-1) and dense block (2-1) are respectively input into transition layer (1-1) and transition layer (2-1) to obtain features of size 37×37 and 19×19. The transition layer includes a convolution layer with a convolution kernel of 1×1 and a stride of 1, and an average pooling layer of 2×2 and a stride of 2.
[0064] S2.6, the obtained 37×37 size features are subsequently passed through dense block (1-2), transition layer (1-2), dense block (1-3), transition layer (1-3), and dense block (1-4); the obtained 19×19 size features are subsequently passed through dense block (2-2), transition layer (2-2), dense block (2-3), transition layer (2-3), and dense block (2-4), among which the number of convolution groups of dense block (1-2), dense block (1-3), and dense block (1-4) are 12, 24, and 16 respectively; the number of convolution groups of dense block (2-2), dense block (2-3), and dense block (2-4) are 12, 24, and 16 respectively, and the transition layer operations are the same; after the above operations, trunk features of size 10×10 and branch features of size 5×5 will be obtained respectively.
[0065] S2.7, the features obtained by the backbone network are subjected to a 3×3 convolution with a stride of 2 to output 5×5 features;
[0066] S2.8, after unifying the feature map sizes of the backbone network and the branch network to 5×5, the number of feature channels is superimposed to obtain 2048 channels, and the features of size 1×1 are obtained through the global average pooling layer;
[0067] S2.9, the obtained features are passed through the classification layer to calculate the classification results, and the total loss of the network is calculated; among them, the classification layer includes a classification convolution layer and a softmax function, and the classification layer is a convolution with a step size of 1 and 10 categories.
[0068] S3: The input of the backbone network, transition layer (1-1) and transition layer (1-2) are halved in size through 3×3 convolutions with a step size of 2, and are superimposed with the output of dense block (2-1), dense block (2-2) and dense block (2-3) of the branch network respectively. Moreover, the features after the maximum pooling layer are halved in size through multiple 3×3 convolutions, and are channel-superimposed with the features after the dense blocks of the branch network to retain the most original feature information.
[0069] Among them, S3 specifically includes the following steps:
[0070] S3.1, the 74×74 size feature obtained in S2.3 is used as input, and the input feature is convolved three times with a 3×3 and a stride of 2 to generate features of sizes 37×37, 19×19, and 10×10 respectively;
[0071] S3.2, according to S3.1, superimpose the features of size 19×19 with the number of output channels of dense block (2-2), and superimpose the features of size 10×10 with the number of output channels of dense block (2-3);
[0072] S3.3, according to S2.3, S2.4, S2.5 and S2.6, the output size of the maximum pooling layer is 74×74, which becomes 37×37 after 3×3 convolution with a step size of 2, and is superimposed with the output channel number of dense block (2-1). The output of the transition layer (1-1) undergoes a 3×3 convolution with a step size of 2 to obtain a feature size of 19×19, which is superimposed with the output channel number of dense block (2-2). The output of the transition layer (1-2) undergoes a 3×3 convolution with a step size of 2 to obtain a feature size of 10×10, which is superimposed with the output channel number of dense block (2-3).
[0073] S4: The Canny algorithm is used to extract the gradient information of the original image, and the gradient enhancement operation is performed with the features after the first dense block of the main line and branch network; moreover, the contour features are enhanced by feature map gradient after the fourth dense block of the main line and branch network.
[0074] Wherein, S4 specifically includes the following steps:
[0075] S4.1, such as Figure 1 As shown by the dotted line in the middle, the contour features of the original image input are calculated using the cloud image input and the Canny edge contour detection operator, which is then converted into feature weights by adding 1 through softmax calculation and multiplied by the outputs of dense blocks (1-1) and dense blocks (2-1) after the S3.3 operation.
[0076] S4.2, such as Figure 1 As shown by the dotted line in the middle, the outputs of dense blocks (1-4) and dense blocks (2-4) are used to calculate the contour features using the Canny edge contour detection operator, and then converted into feature weights by adding 1 through softmax calculation. The weights are multiplied by the outputs of dense blocks (1-4) and dense blocks (2-4) as the final output features.
[0077] Table 1
[0078]
[0079] The present invention not only performs multi-scale feature fusion and uses gradient information, but also superimposes the main line network and the branch network at corresponding size positions in the middle of the model to achieve feature enhancement.
[0080] The model adopted by the method for fusing multi-scale cloud image features and gradient information in the present technical solution is not limited to the DenseNet model, and it can be applied to any model.
[0081] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. 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 cloud image recognition method based on multi-scale features and gradient fusion operation, It is characterized in that include: 1) The feature size of the cloud image is halved by branching, and the multi-scale feature information of the cloud image is increased, including the following steps: After the cloud image passes through a 7×7 convolution layer and a maximum pooling layer, a 3×3 convolution is used to halve the feature map size, and then it is sent to the branch network for training. Finally, the main network and the branch network feature map corresponding to the size position are fused. 2) Enhance the contour information of cloud image features through the gradient feature fusion module, including the following steps: The edge contour extraction algorithm Canny is used to extract gradient information from the feature map after each dense block, and the gradient information is converted into a weight matrix through softmax, and finally a dot multiplication operation is performed with the feature map; The cloud image recognition method based on multi-scale features and gradient fusion operation specifically includes the following steps: S1: Obtain the type image data of the cloud image, divide the cloud image into N categories according to the type characteristics of the cloud image, convert the cloud image size into 294×294, and normalize the image; S2: Input the cloud map into the network to extract image features, and obtain a 64-dimensional feature map of size 74×74 through convolution and maximum pooling layers. The network is divided into a backbone network and a branch network. The backbone network passes through dense block 1-1, transition layer 1-1, dense block 1-2, transition layer 1-2, dense block 1-3, transition layer 1-3 and dense block 1-4 respectively, and the branch network passes through dense block 2-1, transition layer 2-1, dense block 2-2, transition layer 2-2, dense block 2-3, transition layer 2-3 and dense block 2-4 respectively; S3: The input of the backbone network, transition layer 1-1 and transition layer 1-2 are halved by 3×3 convolutions with a step size of 2, and then superimposed with the output of dense block 2-1, dense block 2-2 and dense block 2-3 of the branch network respectively. The features after the maximum pooling layer are halved by multiple 3×3 convolutions, and then channel-superimposed with the features after the dense blocks of the branch network to retain the most original feature information; S4: The Canny algorithm is used to extract the gradient information of the original image, and the gradient enhancement operation is performed with the features after the first dense block of the main line and branch network; the feature map gradient is used to enhance the contour features after the fourth dense block of the main line and branch network.
2. According to claim 1, a cloud image recognition method based on multi-scale features and gradient fusion operation, It is characterized in that The feature map size of the backbone network is 1 times larger than that of the branch network at the same position.
3. According to claim 1, a cloud image recognition method based on multi-scale features and gradient fusion operation, It is characterized in that S1 specifically includes the following steps: S1.1: According to the different characteristics of cloud types, cloud images are divided into N categories; S1.2: The image data pixels are resized to 294×294, and the cloud features are normalized to [-1, 1].
4. According to claim 1, a cloud image recognition method based on multi-scale features and gradient fusion operation, It is characterized in that S2 specifically includes the following steps: S2.1: Input normalized cloud feature map; S2.2: The input image first passes through a convolution layer with a kernel of 7×7 and a stride of 2 to obtain a feature map of size 147×147. Batch normalization and ReLU activation are performed after each convolution layer. S2.3: The obtained features are passed through the maximum pooling layer to obtain a feature map of size 74×74; S2.4, the obtained features are input into the main network and the branch network respectively; the main network is input into the dense block 1-1 to obtain features of size 74×74; the branch network is subjected to a 3×3 convolution operation with a step size of 2 to obtain features of size 37×37, and then input into the dense block 2-1 to obtain features of size 37×37; the features obtained after each group of convolutions will be superimposed with the previously input features by the number of channels as the input for the next convolution; S2.5: The features obtained from dense blocks 1-1 and dense blocks 2-1 are respectively input into transition layer 1-1 and transition layer 2-1 to obtain features of size 37×37 and 19×19; S2.6: The obtained 37×37 size features are subsequently passed through dense block 1-2, transition layer 1-2, dense block 1-3, transition layer 1-3, and dense block 1-4; the obtained 19×19 size features are subsequently passed through dense block 2-2, transition layer 2-2, dense block 2-3, transition layer 2-3, and dense block 2-4; among them, the number of convolution groups of dense block 1-2, dense block 1-3, and dense block 1-4 are 12, 24, and 16 respectively, and the number of convolution groups of dense block 2-2, dense block 2-3, and dense block 2-4 are 12, 24, and 16 respectively; after the above operations, the main features of size 10×10 and the branch features of size 5×5 are obtained respectively; S2.7: The features obtained by the backbone network are subjected to a 3×3 convolution with a stride of 2 to output 5×5 features; S2.8, after unifying the feature map sizes of the backbone network and the branch network to 5×5, the number of feature channels is superimposed to obtain 2048 channels, and the features of size 1×1 are obtained through the global average pooling layer; S2.9, the obtained features are passed through the classification layer to calculate the classification results, and the total loss of the network is calculated; among them, the classification layer includes a classification convolution layer and a softmax function, and the classification layer is a convolution with a step size of 1 and 10 categories.
5. According to claim 4, a cloud image recognition method based on multi-scale features and gradient fusion operation, It is characterized in that The maximum pooling layer described in S2.3 uses a convolution kernel size of 3×3 and a stride of 2.
6. According to claim 4, a cloud image recognition method based on multi-scale features and gradient fusion operation, It is characterized in that In S2.4, dense block 1-1 and dense block 2-1 each contain 6 groups of identical convolutions, each group of convolutions contains 1 convolution layer with a convolution kernel of 1×1 and 1 convolution layer with a convolution kernel of 3×3, and the step size is 1.
7. The cloud image recognition method based on multi-scale feature and gradient fusion operation according to claim 4, It is characterized in that The transition layer in S2.5 includes 1 convolution layer with a convolution kernel of 1×1 and a stride of 1, and 1 average pooling layer of 2×2 and a stride of 2.
8. According to claim 4, a cloud image recognition method based on multi-scale features and gradient fusion operation, It is characterized in that S3 specifically includes the following steps: S3.1: The 74×74 size feature obtained in S2.3 is used as input. The input feature undergoes three 3×3 convolutions with a stride of 2 to generate features of sizes 37×37, 19×19, and 10×10 respectively. S3.2: Overlay the features of size 19×19 with the number of output channels of dense block 2-2, and overlay the features of size 10×10 with the number of output channels of dense block 2-3; S3.3: The output size of the maximum pooling layer is 74×74, which becomes 37×37 after 3×3 convolution with a stride of 2, and is superimposed with the output channel number of dense block 2-1. The output of the transition layer 1-1 undergoes 3×3 convolution with a stride of 2 to obtain a feature size of 19×19, which is superimposed with the output channel number of dense block 2-2. The output of the transition layer 1-2 undergoes 3×3 convolution with a stride of 2 to obtain a feature size of 10×10, which is superimposed with the output channel number of dense block 2-3.
9. The cloud image recognition method based on multi-scale feature and gradient fusion operation according to claim 4, It is characterized in that S4 specifically includes the following steps: S4.1: Using the cloud image input, the Canny edge contour detection operator is used to calculate the contour features of the original image input, and then the softmax calculation is added with 1 to convert it into a feature weight, which is multiplied by the output of dense block 1-1 and dense block 2-1 after S3.3 operation respectively; S4.2: After calculating the contour features of the outputs of dense blocks 1-4 and dense blocks 2-4 using the Canny edge contour detection operator, the softmax calculation is added with 1 to convert them into feature weights, which are multiplied by the outputs of dense blocks 1-4 and dense blocks 2-4 as the final output features.
