CN-121982562-A - SAR image rapid ship detection method capable of improving RT-DETR

CN121982562ACN 121982562 ACN121982562 ACN 121982562ACN-121982562-A

Abstract

The invention discloses an improvement The SAR image rapid ship detection method comprises the steps of S1, obtaining an original ship SAR image, carrying out target annotation on the original ship SAR image to obtain a ship image sample after corresponding target annotation, and S2, improving the ship image sample And S3, carrying out model training on the image rapid ship detection model according to the ship image sample to obtain an optimal detection model, and realizing rapid detection of the ship in the SAR image according to the optimal detection model. The method solves the technical problem that the detection speed of the ship target in the SAR image can not be improved while the detection precision is ensured by the traditional method.

Inventors

ZHOU HUI
HOU CUNLIN
LIU ZHEN

Assignees

大连东软信息学院

Dates

Publication Date: 20260505
Application Date: 20260113

Claims (5)

1. Improved structure The SAR image rapid ship detection method is characterized by comprising the following steps of: S1, acquiring an original ship SAR image and carrying out target labeling on the original ship SAR image to obtain a ship image sample after corresponding target labeling; s2, by improvement The network model is used for constructing an image rapid ship detection model, which comprises the step of improving the original based on an Axial-Attention module Optimized backbone network constructed by backbone network of (a), and improved primitive based on EAA module A lightweight feature fusion module constructed by the encoder of (a) and a decoder; The optimized backbone network is used for enhancing the global receptive field and controlling the corresponding parameter quantity in the process of extracting the multi-resolution characteristic in the ship image sample so as to obtain a multi-scale characteristic map; the lightweight characteristic fusion module is used for realizing characteristic interaction fusion of the multi-scale characteristic images and obtaining lightweight fusion characteristic images; the decoder is used for realizing ship detection in the ship image sample according to the lightweight fusion feature map; And S3, carrying out model training on the image rapid ship detection model according to the ship image sample to obtain an optimal detection model, and realizing rapid detection of the ship in the SAR image according to the optimal detection model.
2. An improvement according to claim 1 The SAR image rapid ship detection method is characterized in that an optimized backbone network constructed in the S2 comprises a first convolution layer, a first stage module, a second stage module, a third stage module and a fourth stage module which are connected in sequence; The first convolution layer is used for carrying out resolution characteristic extraction on an input ship image sample to obtain a first characteristic diagram; The first stage module, the second stage module, the third stage module and the fourth stage module are respectively composed of Bottleneck modules and an Axial-Attention module which are sequentially connected and have the same structure and different convolution kernels, wherein the Bottleneck module is used for carrying out convolution operation on an input feature map; The second convolution layer is used for executing convolution operation on the output of the Bottleneck module, the high-axial multi-head attention layer is used for extracting global correlation characteristics of a feature map of the output of the second convolution layer along the height direction, the first splicing layer is used for executing splicing operation on the output of the high-axial multi-head attention layer, the width axial multi-head attention layer is used for extracting global correlation characteristics of the feature map of the output of the first splicing layer along the horizontal direction, the second splicing layer is used for executing splicing operation on the output of the width axial multi-head attention layer, and the fusion layer is used for executing element-by-element summation operation on the output of the second splicing layer and the output of the Bottleneck module; the first stage module is used for acquiring a second feature map according to the first feature map; the second stage module is used for acquiring a third feature map according to the second feature map; the third stage module is used for acquiring a fourth feature map according to the third feature map; The fourth stage module is used for obtaining a fifth feature map according to the fourth feature map; the multi-scale feature map is a third feature map, a fourth feature map and a fifth feature map.
3. An improvement according to claim 2 The SAR image rapid ship detection method is characterized in that a lightweight characteristic Fusion module constructed in the S2 comprises a fourth convolution layer, a fifth convolution layer, a sixth convolution layer, a flat layer, an EAA module, a seventh convolution layer, a first upsampling layer, a first Fusion layer, an eighth convolution layer, a second upsampling layer, a second Fusion layer, a first downsampling layer, a third Fusion layer, a second downsampling layer and a third Fusion layer; the fourth convolution layer is used for performing convolution operation on the third feature map; The fifth convolution layer is used for performing convolution operation on the fourth feature map; the sixth convolution layer is used for performing convolution operation on the fifth feature map; The EAA module is used for acquiring a global context feature vector according to the output of the flatten layer based on a Transform self-attention mechanism, the seventh convolution layer is used for performing convolution operation on the global context feature vector, the first upsampling layer is used for performing upsampling operation on the output of the seventh convolution layer, and the first Fusion layer is used for performing feature Fusion operation on the output of the first upsampling layer and the output of the fifth convolution layer; the eighth convolution layer is used for performing convolution operation on the output of the first Fusion layer; The second up-sampling layer is used for performing up-sampling operation on the output of the eighth convolution layer; The second Fusion layer is used for executing feature Fusion operation on the output of the second upsampling layer and the output of the fourth convolution layer, the first downsampling layer is used for executing downsampling operation on the output of the second Fusion layer, the third Fusion layer is used for executing feature Fusion operation on the output of the first downsampling layer and the output of the eighth convolution layer, the second downsampling layer is used for executing downsampling operation on the output of the third Fusion layer, and the fourth Fusion layer is used for executing feature Fusion operation on the output of the second downsampling layer and the output of the seventh convolution layer; the lightweight Fusion feature map is Fusion features with different resolutions formed by the output of the second Fusion layer, the output of the third Fusion layer and the output of the fourth Fusion layer.
4. An improvement according to claim 3 The SAR image rapid ship detection method is characterized by comprising the following steps of: S100, acquiring a query matrix And key matrix ; S101, according to the query matrix And parameter vectors that can be learned Acquiring a global attention query vector; and obtain a global attention query vector The expression of (2) is: wherein Representing a feature dimension; S102, query vector for global attention Performing average pooling operation to obtain global query vector ; S103, the global query vector is processed And key matrix And multiplying to obtain the context feature vector.
5. An improvement according to claim 4 The SAR image rapid ship detection method is characterized in that the method for acquiring the optimal detection model in S3 specifically comprises the following steps: S31, randomly dividing a ship image sample into a training set and a verification set according to a preset proportion; S32, performing model training on the image rapid ship detection model according to the training set to obtain a trained image rapid ship detection model; S33, performing model verification on the trained image rapid ship detection model through the verification set based on a model loss function, wherein the model loss function comprises any one of a cross entropy loss function or a mean square error function; Judging whether the output of the trained image rapid ship detection model is converged or not; If the output of the trained image rapid ship detection model converges, confirming that the trained image rapid ship detection model is the optimal detection model; Otherwise, based on the back propagation method, the weight parameters of the trained image rapid ship detection model are adaptively adjusted, and the step S32 is repeatedly executed until the weight parameters of the converged trained image rapid ship detection model are confirmed to be the optimal weight parameters, and the image rapid ship detection model is reconstructed to obtain the optimal detection model.

Description

SAR image rapid ship detection method capable of improving RT-DETR Technical Field The invention relates to the technical field of target detection, in particular to an improvementThe SAR image rapid ship detection method. Background In recent years, more and more students are turning to a target detection algorithm for deep learning. The method utilizes a deep convolution network to automatically extract the characteristics, and replaces the traditional manual characteristics (such as SIFT, HOG and the like). For example, zhao et al developed an embedded ship detection system, applied a deep learning algorithm to ship identification in an actual scene, DONG et al proposed a remote sensing ship image detection method based on an attention mechanism, and improved the target detection capability through a spatial attention mechanism, HAN et al designed a multi-scale feature extraction module, and enhanced the characterization capability of the model to image features, thereby further improving the accuracy, NIU et al improved the convolution structure in YOLO v8 frame, and improved the feature extraction accuracy to irregular shape ship targets. Subsequently, researchers have attempted to introduce transgenes into the field of computer vision due to their high performance and the nature of self-attentive mechanisms. Carion proposes DETR in 2020, which lays a foundation for the application of a transducer in a target detection task. Thereafter, a variety of DETR series algorithms are applied for target detection and exhibit good performance in remote sensing images. Ma et al first applied DETR to directional optical remote sensing target detection. Dai et al further validated the effectiveness of being able to deform DETR in optical images. Although the conventional method and the deep learning method make many developments in ship detection, most of researches focus on improvement of detection accuracy, but neglect optimization of detection speed. This trend can be derived from the actual working environment, since SAR images often take several minutes to hours from acquisition to preprocessing, increasing small amounts of computation time is generally considered acceptable, in contrast to the more significant impact of detection accuracy and accuracy on system performance. However, in scenes with extremely high real-time requirements, such as offshore emergency rescue and military emergency decisions, the detection speed has a crucial meaning. Therefore, improving the detection speed of the ship target in the SAR image while ensuring the detection precision has become one of the key technical challenges in the current target detection field. Disclosure of Invention The present invention provides an improvementThe SAR image rapid ship detection method aims to overcome the technical problems. In order to achieve the above object, the technical scheme of the present invention is as follows: improved structure The SAR image rapid ship detection method specifically comprises the following steps: S1, acquiring an original ship SAR image and carrying out target labeling on the original ship SAR image to obtain a ship image sample after corresponding target labeling; s2, by improvement The network model is used for constructing an image rapid ship detection model, which comprises the step of improving the original based on an Axial-Attention moduleOptimized backbone network constructed by backbone network of (a), and improved primitive based on EAA moduleA lightweight feature fusion module constructed by the encoder of (a) and a decoder; The optimized backbone network is used for enhancing the global receptive field and controlling the corresponding parameter quantity in the process of extracting the multi-resolution characteristic in the ship image sample so as to obtain a multi-scale characteristic map; the lightweight characteristic fusion module is used for realizing characteristic interaction fusion of the multi-scale characteristic images and obtaining lightweight fusion characteristic images; the decoder is used for realizing ship detection in the ship image sample according to the lightweight fusion feature map; And S3, carrying out model training on the image rapid ship detection model according to the ship image sample to obtain an optimal detection model, and realizing rapid detection of the ship in the SAR image according to the optimal detection model. Further, the optimized backbone network constructed in the S2 comprises a first convolution layer, a first stage module, a second stage module, a third stage module and a fourth stage module which are sequentially connected; The first convolution layer is used for carrying out resolution characteristic extraction on an input ship image sample to obtain a first characteristic diagram; The first stage module, the second stage module, the third stage module and the fourth stage module are respectively composed of Bottleneck modules and an Axial-Attention mod