Ship detection method based on reconstruction convolution and enhanced feature fusion
By adopting the feature fusion method of reconstructed convolution and NL-Block modules in ship detection, complex background and multi-scale problems are solved, and the accuracy and robustness of the detection are improved, which is suitable for the deployment of resource-constrained environments.
Patent Information
- Application Number
- CN202410980348.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-12
AI Technical Summary
Existing ship detection algorithms are difficult to accurately focus on ship targets in complex backgrounds, and they face multi-scale problems in ships, resulting in frequent missed inspections and missed inspections.
Using a method based on reconstruction convolution and enhanced feature fusion, the feature extraction backbone network acquires different scale features, performs spatial and channel reconstruction convolution, combines the NL-Block module to fusion high-level semantic information with underlying features, and combines the feature map Enter a one-stage detection algorithm for target box prediction and classification.
Improve the accuracy and robustness of ship detection, reduce model parameters and calculation complexity, and is suitable for deployment in resource-constrained environments, which can better capture ship target characteristics at different scales and enhance the ability to detect multi-scale .
Smart Images

Figure CN118942052A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of ship target detection, and in particular to a ship detection method based on reconstructed convolution and enhanced feature fusion. Background Art
[0002] With the development of information technology, real-time detection and intelligent tracking of ships have been widely used in the field of maritime supervision, greatly improving the efficiency of maritime supervision. However, traditional detection methods generally have problems such as slow training speed and inability to handle ship targets in complex backgrounds, making it difficult to achieve high-precision detection and tracking of ships. In recent years, with the development of deep learning, the use of convolutional neural network models to achieve ship detection has become the current mainstream ship detection method. Although deep learning-based ship detection methods have made breakthrough progress, due to the diverse scales of ships, changeable climate environments, and complex port backgrounds, existing ship detection algorithms still have many missed detections and false detections. How to improve the robustness of ship detection models is currently a hot topic and difficulty in ship detection research.
[0003] At present, the target detection algorithms based on deep learning are mainly divided into two categories:
[0004] (1) Two-stage object detection algorithm. In the first stage, candidate boxes that may contain objects are generated through the Region Proposal Network (RPN) or other methods. These candidate boxes represent potential object areas in the image. In the second stage, these candidate boxes are sent to the classifier for object classification and position fine-tuning, and finally the object detection results are obtained. Typical two-stage object detection algorithms include R-CNN and Mask R-CNN.
[0005] (2) The one-stage target detection algorithm directly processes the entire image to complete the target classification and positioning tasks. This method is simple and efficient, and can perform target detection in real time. It is particularly suitable for application scenarios that require detection speed. Representative one-stage target detection algorithms include SSD and YOLO. Although in some cases, the one-stage algorithm may sacrifice a certain degree of accuracy, it is widely used in real-time target detection tasks due to its fast detection speed.
[0006] The one-stage ship detection method based on deep learning performs better than the two-stage method in dealing with complex scenarios, providing a reliable solution for maritime supervision. However, due to the diversity of ship scales, the variability of climate and environment, and the complexity of port backgrounds, ship detection still faces the following challenges:
[0007] 1. It is difficult for the model to accurately focus on the ship target in a complex background, causing the ship feature information to be disturbed and become unclear, which in turn affects the accurate positioning and classification of the ship. Traditional ship detection algorithms usually only use convolutional neural networks as the backbone feature extraction network for deep feature extraction. However, this approach may cause the network to have spatial and channel feature redundancy, reducing the performance and efficiency of the model.
[0008] 2. Ship detection faces the problem of multi-scale ships. To address this problem, previous ship detection models usually use feature pyramid fusion technology, but simple fusion methods may lead to insufficient fusion of deep features and shallow features, and fail to effectively utilize information from features at different levels, resulting in missed detections in multi-scale ship detection. Summary of the invention
[0009] The purpose of the present invention is to overcome the problems that the models in the above-mentioned prior art are difficult to accurately focus on ship targets in complex backgrounds and that ship detection faces problems such as multi-scale ships, and to provide a ship detection method based on reconstructed convolution and enhanced feature fusion.
[0010] The purpose of the present invention can be achieved by the following technical solutions:
[0011] A ship detection method based on reconstruction convolution and enhanced feature fusion, the method steps comprising:
[0012] Acquire images containing ship targets;
[0013] The feature extraction backbone network is used to extract features from images containing ship targets, and features of three different scales are obtained;
[0014] The extracted features are convolved for spatial and channel reconstruction to eliminate spatial redundancy and channel redundancy, and three feature layers are obtained;
[0015] The three feature layers are input into the NL-Block module to fuse the high-level semantic information with the low-level feature information, and three feature maps are output;
[0016] The feature map is sent to the detection head of the one-stage detection algorithm to predict and classify the target box to obtain the ship detection result.
[0017] As a preferred technical solution, the space and channel reconstruction convolution retains important spatial information and reduces redundancy between channels by performing feature separation and reconstruction operations on space and channels respectively.
[0018] As a preferred technical solution, the specific steps for the spatial reconstruction are as follows:
[0019] Perform group normalization on the input feature map;
[0020] After the normalization operation, the feature layer is judged by setting the threshold, and the feature layer W with rich spatial information is 1 With less spatial information feature layer W 2 separation;
[0021] The separated feature layers are reconstructed and the feature layers with rich spatial information are divided into W 11 , W 12 The feature layer with less spatial information is divided into two parts: 21 , W 22 The two parts are cross-reconstructed to obtain the spatial refinement feature map X through the following formula w :
[0022] W 11 +W 22 =X w1
[0023] W 12 +W 21 =X w2
[0024] X w1 ∪X w2 =X w .
[0025] As a preferred technical solution, the specific steps for channel reconstruction are as follows:
[0026] Use the hyperparameter α to reconstruct the space to obtain the spatial refinement feature map X w It is divided into two parts. One part has a channel number of αC, which is used as the input X for rich feature extraction. up , the other part takes (α-1)C as a supplement to the rich feature extraction X low ;
[0027] Input X for rich feature extraction up Perform group convolution and point-by-point convolution respectively, and add the results to get Y 1 , a supplement to rich feature extraction low Perform point-by-point convolution to get Y 2 ;
[0028] Perform channel fusion operation, combine global spatial information and channel statistical information through global average pooling technology, and obtain feature weight vector β after Softmax processing 1 , β 2 , output Y = β 1 Y 1 +β 2 Y 2 .
[0029] As a preferred technical solution, the feature fusion of high-level semantic information and low-level feature information is specifically as follows:
[0030] The three feature layers are shallow features, middle features, and deep features.
[0031] The NL-Block module is used to enhance the fusion of shallow features and deep features as well as middle features and deep features, and finally three feature maps of the fused features and deep features are output.
[0032] As a preferred technical solution, the NL-Block module combines features containing fine-grained pixel information with features containing a larger receptive field by calculating the global dependency of high-level features and low-level features. The specific steps are as follows:
[0033] Adjust the high-level feature map to the same size as the low-level feature map to ensure that the high-level features are aligned with the low-level features in the spatial dimension;
[0034] Pull the adjusted high-level feature map and low-level feature map into one-dimensional vectors respectively, and retain the feature information of each position in the feature map;
[0035] The similarity between the high-level feature map and the low-level feature map is used to measure the correlation between the low-level features and the high-level features at each position;
[0036] According to the similarity weight, each position of the high-level feature map is weighted summed to obtain the fused feature representation.
[0037] As a preferred technical solution, during the one-stage detection algorithm training process, the total loss function of the model adopts the classification loss L class , bounding box regression loss L local And the confidence loss L obj The weighted sum of is calculated as follows:
[0038]
[0039] Among them, λ is the weight of the regression box loss, which is used to adjust the model's attention to the regression box loss.
[0040] As a preferred technical solution, the classification loss L class The calculation formula is as follows:
[0041]
[0042] Where N represents the total number of categories; x i is the predicted value of the current category; y i is the probability of the current category; is the true value of the current category.
[0043] As a preferred technical solution, the border regression loss L local Using CIOU, the calculation formula is as follows:
[0044]
[0045] Among them, ν is the similarity of the aspect ratio between the real box and the predicted box; α is the balance factor, which is used to adjust the weight of the aspect ratio when calculating the loss; IOU is the intersection-union ratio of the target box.
[0046] As a preferred technical solution, the confidence loss L obj The calculation formula is as follows:
[0047]
[0048] C * =Sigmod(C)
[0049] Where N is the number of samples; γ is the balance factor, which is used to balance the impact of positive and negative samples on the loss; It is the confidence of the predicted value C after passing through the Sigmod function.
[0050] Compared with the prior art, the present invention has the following beneficial effects:
[0051] 1) In the ship detection task, the use of the reconstructed convolution SCConv module brings new ideas to the design of convolutional neural networks. Traditional convolutional neural networks face the problem of detecting multi-scale targets when dealing with ship detection tasks. It is necessary to minimize the complexity and computational complexity of the model while ensuring the detection accuracy. The introduction of the SCConv module can effectively reduce the number of parameters and computational complexity of the model while maintaining the performance of the model. By embedding the SCConv module into the convolutional neural network, it can replace the traditional convolutional layer to achieve better feature extraction effects. In the ship detection task, the SCConv module can help the network better capture the features of ship targets of different scales and improve detection accuracy and robustness. In addition, due to the lightweight and high efficiency of the SCConv module, the ship detection model can still obtain excellent detection performance in a resource-constrained environment, so it is suitable for deployment on embedded devices or mobile devices.
[0052] 2) The present invention utilizes the fusion capability of NL-Block to effectively solve the multi-scale problem in the ship detection task. In ship detection, the size and shape of the ship are affected by the distance and angle, resulting in ship targets of different scales appearing in the same image. Traditional detection methods cannot effectively cope with such multi-scale changes. By combining shallow features (with high spatial resolution) with deep features (with large receptive field), NL-Block can help the model better capture the characteristics of ship targets at different scales. Shallow features provide details and local information, which helps to accurately locate small-scale ships, while deep features provide global semantic information, which helps to identify large-scale ships. The NL-Block model can more comprehensively understand the ship targets in the image and improve the detection accuracy and robustness of ships of different scales. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 It is a flow chart of the implementation of the ship detection method based on reconstruction convolution and enhanced feature fusion of the present invention;
[0054] Figure 2 It is a schematic diagram of the architecture of the ship detection method based on reconstruction convolution and enhanced feature fusion of the present invention;
[0055] Figure 3 Schematic diagram of SCConv calculation process;
[0056] Figure 4 It is a schematic diagram of the SRU calculation process;
[0057] Figure 5 This is a schematic diagram of the CRU calculation process;
[0058] Figure 6 This is a schematic diagram of the NL-Block calculation process;
[0059] Figure 7 The figure is a flow chart of a ship detection method based on reconstruction convolution and enhanced feature fusion in an embodiment of the present invention. DETAILED DESCRIPTION
[0060] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.
[0061] Example 1
[0062] The present invention proposes a ship detection method based on reconstruction convolution and enhanced feature fusion to obtain better ship detection results, so that it has good performance when dealing with complex scenes and multi-scale ship problems. First, reconstruction convolution is introduced to eliminate spatial redundancy and channel redundancy that may be generated after the convolution operation. Reconstruction convolution can effectively reduce the number of parameters and calculations while keeping the size of the feature map unchanged, making the model more lightweight. In addition, the present invention proposes an enhanced feature fusion method called NL-Block. NL-Block uses the mechanism of non-local blocks to capture the global dependency between high-level features and shallow features. NL-Block can more fully integrate high-level semantic information and underlying feature information, thereby improving the model's ability to detect multi-scale ships. Figure 1 As shown, the specific detection process of the ship detection method proposed by the present invention is as follows:
[0063] S1. Take an image containing a ship target as input data.
[0064] S2. Extract features from the input data through a feature extraction backbone network to obtain features of three different scales.
[0065] S3. Reconstruct the convolution of the extracted features to reduce the redundancy of spatial and channel information and obtain three feature layers of different sizes.
[0066] S4: Input three feature layers of different sizes into the NL-Block module and perform feature fusion on the high-level semantic information and the low-level feature information, and output three feature maps.
[0067] The present invention is based on the architecture of the reconstruction convolution and enhanced feature fusion method. Figure 2 shown.
[0068] S5. Send the three output feature maps to the detection head of the one-stage detection algorithm for target box prediction and classification.
[0069] 1. The present invention uses reconstruction convolution to eliminate spatial redundancy and channel redundancy of the convolutional neural network model. The spatial and channel reconstruction convolution (SCConv) module aims to reduce the information redundancy between feature maps by optimizing the feature extraction process of the convolutional neural network, thereby compressing the model and improving performance. Its core idea is to retain important spatial information and reduce redundancy between channels through feature separation and reconstruction operations, thereby reducing model parameters and computational complexity while improving the model's representation ability. The calculation process of SCConv is as follows: Figure 3 As shown. The implementation of SCConv includes the following two key steps:
[0070] (1.1) Spatial Reconstruction (SRU): First, a feature separation operation is performed, the purpose of which is to separate feature maps with large information content from feature maps with small information content. The formula is as follows, where X is the original input feature map:
[0071]
[0072] W=Gate(Sigmoid(W γ (GN(X))))
[0073] First, the input feature map is grouped and normalized: out =GN(X), where γ and β are two training variables. γ represents the spatial information between pixels. The larger its value, the richer the spatial information. Normalization will change the distribution characteristics of the input and may lose some valuable information. β, as a learnable offset parameter, can help the network recover the expression ability lost during the normalization process.
[0074] After the normalization operation, the feature layer is judged by setting a threshold. The feature layer above the threshold is called the feature layer with rich spatial information, and the feature layer below the threshold is the feature layer with less spatial information. The feature layer with rich spatial information is separated from the feature layer with less spatial information by the above formula. The feature layer with rich spatial information is W 1 , the feature layer with less spatial information is W 2 .
[0075] Then the reconstruction operation is performed, which adds the features with more information and the features with less information to generate more informative features and save space. The calculation process is: divide the feature layer with rich spatial information into W 11 , W 12 The feature layer with less spatial information is divided into two parts: 21 , W 22 Two parts, and cross reconstruction is performed by the following formula:
[0076] W 11 +W 22 =X w1
[0077] W 12 +W 21 =X w2
[0078] X w1 ∪X w2 =X w
[0079] After SRU processing, the spatial refinement feature map X can be obtained w, which separates features with large amounts of information from features with small amounts of information, reduces feature redundancy in the spatial dimension, and enables the neural network to better utilize information between different feature maps, improves the representation ability of the model, and can reduce the number of model parameters and computational complexity to a certain extent. The calculation process of SRU is as follows Figure 4 shown.
[0080] (1.2) Channel reconstruction (CRU): Channel reconstruction also performs separation followed by reconstruction. First, we use the hyperparameter α (0≤α≤1) to transform X w It is divided into two parts. One part has a channel number of αC and is recorded as the input X of "rich feature extraction". up , the other part is (α-1)C, recorded as the supplementary X of "rich feature extraction" low . up Perform group convolution and point-by-point convolution respectively, and add the results to get Y 1 , for X low Perform point-by-point convolution to get Y 2 .
[0081] Then the channel fusion operation is performed. This operation first combines the global spatial information and channel statistical information through the global average pooling technology, and then obtains the feature weight vector β after Softmax processing. 1 , β 2 , and finally output Y = β 1 Y 1 +β 2 Y 2 Channel reconstruction can make the neural network use feature information more efficiently and further improve the representation ability of the model. The calculation process of CRU is as follows: Figure 5 shown.
[0082] 2. Use enhanced feature fusion methods to improve the degree of fusion between feature layers. In ship detection tasks, it is also crucial to fuse features of different scales. In actual detection, there may be both large and small ships in the image. If the feature representations of ships of different scales are inconsistent, using the fused features for ship detection will reduce the detection accuracy. Therefore, for ship detection tasks, it is necessary to consider both local and global information in the image. Shallow features can help capture the detailed features and local structures of ships in ship detection, which is very important for accurately detecting small ships. On the other hand, high-level features can better represent the overall shape and size of the ship. By fusing these two features, the model can be made more accurate and robust when detecting ships of various sizes. To this end, the present invention uses the NL-Block (NonLocal-Block) fusion strategy to better fuse shallow features and high-level features, and its calculation process is as follows: Figure 6 shown.
[0083] In NL-Block, we combine features containing fine-grained pixel information with features containing larger receptive fields by calculating the global dependency between high-level features and shallow-level features. The specific calculation steps are as follows:
[0084] (2.1) Feature transformation: Through operations such as convolution or upsampling, the high-level feature map (Key and Value) is adjusted to the same size as the low-level feature map (Query), ensuring that the high-level features are aligned with the low-level features in the spatial dimension.
[0085] (2.2) Pull into one-dimensional vector: Pull the adjusted high-level feature map and low-level feature map into one-dimensional vectors respectively, retaining the feature information of each position in the feature map.
[0086] (2.3) Calculate similarity: Use the similarity between the high-level feature map Key and the low-level feature map Query to measure the degree of correlation between the low-level features and high-level features at each position.
[0087] (2.4) Weighted fusion: According to the similarity weights obtained in step (3), a weighted sum is performed on each position of the high-level feature map to obtain a fused feature representation, thereby achieving modeling of global dependencies.
[0088] Specific implementation examples:
[0089] In the specific implementation process, this embodiment selects YOLOv7 as the backbone network, and uses the reconstructed convolution and enhanced feature fusion method based on the network. Figure 7 As shown in the figure, the experimental data uses the Seaship dataset, which has 7,000 images, including ship images with different backgrounds and scales. The specific execution process is as follows:
[0090] 1) The images in the dataset are used as input data and input into the YOLOv7 backbone feature extraction network for feature extraction.
[0091] 2) The extracted features of three scales, 80x80, 40x40, and 20x20, are reconstructed and convolved to eliminate spatial redundancy and channel redundancy without changing their scale and number of channels.
[0092] 3) Perform feature fusion on the reconstructed features. The specific fusion process is: the three feature layers are used as shallow features, middle features, and deep features respectively, among which the shallow features are enhanced fused with the deep features, and the middle features are enhanced fused with the deep features. Finally, the three fused feature maps are input into the YOLO detection head respectively.
[0093] 4) After receiving three feature layers of different sizes, the YOLO detection head generates a candidate box, and then selects the most suitable target box as the prediction box through the non-maximum suppression algorithm, and classifies the target box at the same time to generate a true box and classification confidence.
[0094] During the training process, the model of the present invention adopts the following three loss functions: classification loss L class , bounding box regression loss L local , confidence loss L obj The total loss function of the model is the weighted sum of the above three types of losses, and its calculation formula is as follows:
[0095]
[0096] Among them, λ is the weight of the regression box loss, which is used to adjust the model's attention to the regression box loss.
[0097] Classification loss L class The calculation formula is as follows:
[0098]
[0099] Where N represents the total number of categories, x i is the predicted value of the current category, y i is the probability of the current category, is the true value of the current category (0 or 1).
[0100] Confidence loss L obj The calculation formula is as follows:
[0101]
[0102] C * =Sigmod(C)
[0103] Among them, N is the number of samples, and γ is the balance factor, which is used to balance the impact of positive and negative samples on the loss. It is the confidence of the predicted value C after passing through the Sigmod function.
[0104] In calculating the bounding box regression loss L local CIOU is used when calculating the following formula:
[0105]
[0106] Among them, v is the similarity of the aspect ratio between the real box and the predicted box, and its calculation formula is as follows:
[0107]
[0108] Among them, α is the balance factor, which is used to adjust the weight of width and height similarity when calculating the loss. IOU is the intersection-over-union ratio of the target box. The corresponding calculation formula is as follows:
[0109]
[0110] The final regression loss L local The calculation formula is:
[0111]
[0112] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.
Claims
1. A ship detection method based on reconstruction convolution and enhanced feature fusion, characterized in that: The method steps include: Acquire images containing ship targets; The feature extraction backbone network is used to extract features from images containing ship targets, and three features of different scales are obtained; The extracted features are convolved for spatial and channel reconstruction to eliminate spatial redundancy and channel redundancy, and three feature layers are obtained; The three feature layers are input into the NL-Block module to fuse the high-level semantic information with the low-level feature information, and three feature maps are output; The feature map is sent to the detection head of the one-stage detection algorithm to predict and classify the target box to obtain the ship detection result.
2. A ship detection method based on reconstruction convolution and enhanced feature fusion according to claim 1, characterized in that: The spatial and channel reconstruction convolutions retain important spatial information and reduce redundancy between channels by performing feature separation and reconstruction operations on space and channels respectively.
3. A ship detection method based on reconstruction convolution and enhanced feature fusion according to claim 2, characterized in that: The specific steps for the spatial reconstruction are as follows: Perform group normalization on the input feature map; After the normalization operation, the feature layer is judged by setting a threshold, and the feature layer W1 with rich spatial information is separated from the feature layer W2 with less spatial information; The separated feature layers are reconstructed and the feature layers with rich spatial information are divided into W 11 , W 12 The feature layer with less spatial information is divided into two parts: 21 , W 22 The two parts are cross-reconstructed to obtain the spatial refinement feature map X through the following formula w : W 11 +W 22 =X w1 W 12 +W 21 =X w2 X w1 ∪X w2 =X w 。 4. A ship detection method based on reconstruction convolution and enhanced feature fusion according to claim 2, characterized in that: The specific steps for channel reconstruction are as follows: Use the hyperparameter α to reconstruct the space to obtain the spatial refinement feature map X w It is divided into two parts. One part has a channel number of αC, which is used as the input X for rich feature extraction. up , the other part takes (α-1)C as a supplement to the rich feature extraction X low ; Input X for rich feature extraction up Perform group convolution and point-by-point convolution respectively, and add the results to get Y1, which is a supplement to the rich feature extraction X low Perform point-by-point convolution to obtain Y2; Perform channel fusion operation, combine global spatial information and channel statistical information through global average pooling technology, obtain feature weight vectors β1, β2 after Softmax processing, and output Y=β1Y1+β2Y2.
5. The ship detection method based on reconstruction convolution and enhanced feature fusion according to claim 1 is characterized in that: The feature fusion of high-level semantic information and low-level feature information is specifically as follows: The three feature layers are shallow features, middle features, and deep features. The NL-Block module is used to enhance the fusion of shallow features and deep features as well as middle features and deep features, and finally three feature maps of the fused features and deep features are output.
6. A ship detection method based on reconstruction convolution and enhanced feature fusion according to claim 5, characterized in that: The NL-Block module combines features containing fine-grained pixel information with features containing larger receptive fields by calculating the global dependency of high-level features and low-level features. The specific steps are as follows: Adjust the high-level feature map to the same size as the low-level feature map to ensure that the high-level features are aligned with the low-level features in the spatial dimension; Pull the adjusted high-level feature map and low-level feature map into one-dimensional vectors respectively, and retain the feature information of each position in the feature map; The similarity between the high-level feature map and the low-level feature map is used to measure the correlation between the low-level features and the high-level features at each position; According to the similarity weight, each position of the high-level feature map is weighted summed to obtain the fused feature representation.
7. A ship detection method based on reconstruction convolution and enhanced feature fusion according to claim 1, characterized in that: During the one-stage detection algorithm training process, the total loss function of the model adopts the classification loss L class , bounding box regression loss L local And the confidence loss L obj The weighted sum of is calculated as follows: Among them, λ is the weight of the regression box loss, which is used to adjust the model's attention to the regression box loss.
8. A ship detection method based on reconstruction convolution and enhanced feature fusion according to claim 7, characterized in that: The classification loss L class The calculation formula is as follows: Where N represents the total number of categories; x i is the predicted value of the current category; y i is the probability of the current category; is the true value of the current category.
9. A ship detection method based on reconstruction convolution and enhanced feature fusion according to claim 7, characterized in that: The bounding box regression loss L local Using CIOU, the calculation formula is as follows: Among them, v is the similarity of the aspect ratio between the real box and the predicted box; α is the balance factor, which is used to adjust the weight of the aspect ratio when calculating the loss; IOU is the intersection-union ratio of the target box.
10. A ship detection method based on reconstruction convolution and enhanced feature fusion according to claim 7, characterized in that: The confidence loss L obj The calculation formula is as follows: C * =Sigmod(C) Where N is the number of samples; γ is the balance factor, which is used to balance the impact of positive and negative samples on the loss; It is the confidence of the predicted value C after passing through the Sigmod function.
Citation Information
Cited By
Infrared ship image classification method, system and equipment based on feature reconstruction
CN119919745A
Adaptive fine-grained ocean ship rescue positioning method based on severe weather
CN119963649A
Adaptive Fine-Grained Ocean Vessel Rescue Positioning Method Based on Severe Weather
CN119963649B
