A scale-mixed search and refinement SAR ship intelligent detection method

Abstract

The present invention discloses a scale-mixed search and refinement SAR ship intelligent detection method, which relates to the technical field of SAR ship target detection. The key points of the technical solution are: S1. Acquire images of different scales and mine more ship target information through a scaling strategy; S2. Acquire different-level features of scale mixing and integrate information of the same level at different scales through a search module to capture the differential information between ship targets and backgrounds; S3. Aggregate and enhance features of scale mixing between different levels through a refinement module to mine key semantic clues for ship detection. A large number of comparative experiments on the SAR Ship Detection Dataset (SSDD) between this method and existing target detection methods show that, whether compared with the baseline method or the existing target detection method, the proposed method has significant improvements in the overall map, the map in the nearshore scene, and the map in the farshore scene. The proposed mixed-scale search and refinement strategy is beneficial to ship target detection.

Description

Technical Field

[0001] The present invention relates to the technical field of SAR ship target detection, and more particularly to a scale-mixed search and refinement SAR ship intelligent detection method. Background Art

[0002] Existing SAR ship target detection methods based on deep learning only perceive ship targets at a single scale, which makes it difficult to fully exploit the differentiated information between ship targets and cluttered backgrounds. Therefore, the reliability of ship target detection and the accuracy of classification recognition are limited.

[0003] Difficulty and significance of solving existing technical problems: Ship detection in SAR images remains challenging due to the inherent speckle noise in SAR images and interference from background objects with strong scattering intensity, which are very similar to real ship targets. Summary of the Invention

[0004] The purpose of the present invention is to solve the above problems. This method adopts a scaling strategy to obtain more discriminative ship semantic clues by exploring and integrating the differentiated information between ship targets and mixed background at different scales, aggregating and enhancing the features between different levels at mixed scales. A scale-mixed search and refinement SAR ship intelligent detection network is proposed, which effectively mines the semantic clues of ship targets and improves the performance of the SAR ship detection algorithm.

[0005] The above technical objectives of the present invention are achieved through the following technical solutions: a scale-mixed search and refinement SAR ship intelligent detection method, comprising the following steps:

[0006] S1. Obtain images of different scales and mine more ship target information through scaling strategies;

[0007] S2, obtains different levels of features at mixed scales, integrates information of the same level at different scales through the search module, and captures the differentiated information between the ship target and the background;

[0008] S3. Through the refinement module, we aggregate and enhance the mixed-scale features at different levels to mine the key semantic clues for ship detection.

[0009] The present invention is further configured as follows: in step S1, the scaling strategy is used to mine more ship target information, and the input original image is regarded as the main scale, which is recorded as I 1.0× The two auxiliary scales are obtained by upsampling and downsampling the main scale, denoted as I 1.5× , I 0.5× .

[0010] The present invention is further configured as follows: in step S2, the search module integrates information of the same level at different scales to fully capture the differential information between the ship target and the background. The specific principle is as follows:

[0011] L1, the three different scale images I obtained in step S1 1.5× , I 0.5× , I 1.0× At the same time, a weight-shared feature encoder is input for feature extraction, and the feature representation f of the corresponding scale is output i 0.5× , f i 1.0× , f i 1.5× ;

[0012] L2, feature representation f of different scales i 0.5× , f i 1.5× Adjust the scales respectively to make them consistent with the main scale f i 1.0× The characteristics remain the same; for f i 1.5× , use the maximum pooling layer and the average pooling layer for downsampling; for f i 0.5× , upsample it using bilinear interpolation. The features are combined using a concatenation operation, which can be expressed as:

[0013] f i =Concat(f i 0.5× , f i 1.0× , f i 1.5× )

[0014] L3, the combined feature f i Aggregation is performed through a global average pooling layer, and the corresponding channel attention weights are obtained through 1D convolution and sigmoid function. The channel attention weights are multiplied by the input features to obtain the final output features. The process can be expressed as:

[0015]

[0016] Where, represents one-dimensional convolution, σ represents the sigmoid activation function, and GAP represents the global average pooling layer.

[0017] The present invention is further configured as follows: in step S3, the features of scale mixing at different levels are aggregated and enhanced by the refinement module to mine key semantic clues for ship detection. The specific principle is as follows:

[0018] M1, divide the mixed-scale features of different levels obtained in the above step S2 into low-level features {f1, f2}, middle-level features {f3}, and high-level features {f4, f5}, and fuse the low-level features and high-level features respectively through the concatenation operation Concat(·). The process can be expressed as:

[0019] f h =Concat(f4,f5)

[0020] f l =Concat(f1,f2);

[0021] M2: Use the fused high-level features to guide the learning of intermediate features, and use the intermediate features to guide the learning of low-level features. The process can be expressed as:

[0022] f hm =Concat(f h , f m )

[0023] f ml =Concat(f hm , f l );

[0024] M3, using the receptive field component, integrates more discriminative feature representations by expanding the receptive field. The process can be expressed as:

[0025] f′ h =RF(f h )

[0026] f′ hm =RF(f hm )

[0027] f′ ml =RF(f ml )

[0028] Where RF(·) represents the receptive field component, f′ h , f′ hm and f′ ml As the final output prediction layer, it is input into the region proposal network to generate candidate regions for potential targets.

[0029] In summary, the present invention has the following beneficial effects: Extensive comparative experiments comparing the proposed method with existing target detection methods on the SAR Ship Detection Dataset (SSDD) show that compared with the baseline method, the proposed method improves the overall map, the map in nearshore scenarios, and the map in farshore scenarios by 4.2%, 3.27%, and 5.18%, respectively, achieving significant performance improvements at a low parameter cost. Compared with existing target detection methods, the proposed method improves the overall map by 0.27% and the map in farshore scenarios by 2.01%, respectively. These results demonstrate that the proposed hybrid-scale search and refinement strategy is beneficial for ship target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 2. Schematic diagram of a scale-mixed search and refinement SAR ship intelligent detection network in an embodiment of the present invention;

[0031] Figure 2 This is a network structure diagram of a search module in an embodiment of the present invention;

[0032] Figure 3 This is a diagram of a network structure of a mesoscale aggregation unit in an embodiment of the present invention;

[0033] Figure 4 is a network structure diagram of a refinement module in an embodiment of the present invention;

[0034] Figure 5 2 is a network structure diagram of the receptive field component in an embodiment of the present invention.

[0035] Figure 6 It is a visualization diagram of the detection results of the model ablation experiment in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described in detail below in conjunction with the embodiments of the present invention and the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not 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 efforts should fall within the scope of protection of the present invention.

[0037] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the embodiments.

[0038] Example:

[0039] like Figures 1-6 As shown, a scale-mixed search and refinement SAR ship intelligent detection method includes the following steps:

[0040] S1. We analyze the problem that existing SAR target detection methods only perceive ship targets at a single scale and cannot fully mine the differentiated information between ship targets and cluttered backgrounds. We use a scaling strategy to mine more ship target information.

[0041] S2. Obtain images of different scales in the above step S1, and integrate the information of the same level at different scales through the search module to capture the differential information between the ship target and the background;

[0042] S3. In step S2 above, the scale-mixed features of different levels are obtained. The scale-mixed features of different levels are aggregated and enhanced through the refinement module to mine the key semantic clues for ship detection.

[0043] In step S1, the scaling strategy is used to mine more ship target information. For the input original image, it is regarded as the main scale, which is recorded as I 1.0× The two auxiliary scales are obtained by upsampling and downsampling the main scale, denoted as I 1.5× , I 0.5× .

[0044] In step S2, the search module integrates information of the same level at different scales to fully capture the differentiated information between the ship target and the background. The specific principles are as follows:

[0045] L1, the three different scale images I obtained in step S1 1.5× , I 0.5× , I 1.0× At the same time, a weight-shared feature encoder is input for feature extraction, and the feature representation f of the corresponding scale is output i 0.5× , f i 1.0× , f i 1.5× ;

[0046] L2, feature representation f of different scales i 0.5× , f i 1.5× Adjust the scales respectively to make them consistent with the main scale f i 1.0× The characteristics remain the same; for f i 1.5× , use the maximum pooling layer and the average pooling layer for downsampling; for f i 0.5× , upsample it using bilinear interpolation. The features are combined using a concatenation operation, which can be expressed as:

[0047] f i=Concat(f i 0.5× , f i 1.0× , f i 1.5× )

[0048] L3, the combined feature f i Aggregation is performed through a global average pooling layer, and the corresponding channel attention weights are obtained through 1D convolution and sigmoid function. The channel attention weights are multiplied by the input features to obtain the final output features. The process can be expressed as:

[0049]

[0050] Where, represents one-dimensional convolution, σ represents the sigmoid activation function, and GAP represents the global average pooling layer.

[0051] In step S3, the refinement module aggregates and enhances the mixed-scale features at different levels to mine key semantic clues for ship detection. The specific principles are as follows:

[0052] M1, divide the mixed-scale features of different levels obtained in the above step S2 into low-level features {f1, f2}, middle-level features {f3}, and high-level features {f4, f5}, and fuse the low-level features and high-level features respectively through the concatenation operation Concat(·). The process can be expressed as:

[0053] f h =Concat(f4,f5)

[0054] f l =Concat(f1,f2);

[0055] M2: Use the fused high-level features to guide the learning of intermediate features, and use the intermediate features to guide the learning of low-level features. The process can be expressed as:

[0056] f hm =Concat(f h , f m )

[0057] f ml =Concat(f hm , f l );

[0058] M3, using the receptive field component, integrates more discriminative feature representations by expanding the receptive field. The process can be expressed as:

[0059] f′ h =RF(fh )

[0060] f′ hm =RF(f hm )

[0061] f′ ml =RF(f ml )

[0062] Where RF(·) represents the receptive field component, f′ h , f′ hm and f′ ml As the final output prediction layer, it is input into the region proposal network to generate candidate regions for potential targets.

[0063] This specific embodiment is merely an explanation of the present invention and is not intended to limit the present invention. After reading this specification, those skilled in the art may make non-creative modifications to this embodiment as needed. However, as long as such modifications are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A scale-mixed search and refinement SAR ship intelligent detection method, characterized by: The following steps are involved: S1. Obtain images of different scales and mine more ship target information through scaling strategies; S2. Obtaining different-level features of mixed scales, integrating information of the same level at different scales through the search module, and capturing differentiated information between ship targets and backgrounds; S3, through the refinement module aggregation and enhancement of mixed features at different levels and scales, to mine key semantic clues for ship detection; In step S1, the scaling strategy is used to mine more ship target information. For the input original image, it is regarded as the main scale, which is recorded as I 1.0× ; The main scale is upsampled and downsampled to obtain two auxiliary scales, denoted as I 1.5 × , I 0.5× ; In step S2, the search module integrates information of the same level at different scales to fully capture the differentiated information between the ship target and the background. The specific principles are as follows: L1, the three different scale images I obtained in step S1 1.5× , I 0.5× , I 1.0× At the same time, a weight-sharing feature encoder is input for feature extraction, and the feature representation of the corresponding scale is output. L2, feature representation of different scales Adjust the scales separately to match the main scale The characteristics remain the same; for Use the max pooling layer and the average pooling layer for downsampling; for It is upsampled using bilinear interpolation; the features are combined using a concatenation operation, which can be expressed as: L3, the combined feature f i Aggregation is performed through a global average pooling layer, and the corresponding channel attention weights are obtained through 1D convolution and sigmoid function. The channel attention weights are multiplied by the input features to obtain the final output features. The process can be expressed as: Where, represents one-dimensional convolution, σ represents the sigmoid activation function, and GAP represents the global average pooling layer; In step S3, the refinement module aggregates and enhances the mixed-scale features at different levels to mine key semantic clues for ship detection. The specific principles are as follows: M1, divide the mixed-scale features of different levels obtained in the above step S2 into low-level features {f1, f2}, middle-level features {f3}, and high-level features {f4, f5}, and fuse the low-level features and high-level features respectively through the concatenation operation Concat(·). The process can be expressed as: f h =Concat(f4,f5) f l =Concat(f1,f2); M2: Use the fused high-level features to guide the learning of intermediate features, and use the intermediate features to guide the learning of low-level features. The process can be expressed as: f hm =Concat(f h ,f m ) f ml =Concat(f hm ,f l ); M3, using the receptive field component, integrates more discriminative feature representations by expanding the receptive field. The process can be expressed as: f′ h =RF(f h ) f′ hm =RF(f hm ) f′ ml =RF(f ml ) Where RF(·) represents the receptive field component, f′ h , f′ hm and f′ ml As the final output prediction layer, it is input into the region proposal network to generate candidate regions for potential targets.

Images