Research On Ship Detection In Optical Remote Sensing Image Based On Lightweight CNN

Accurate detection of ships is of great significance to the safety of marine traffic.With the rapid development of deep learning,ship detection from optical satellite remote sensing images using Convolutional Neural Network(CNN)can significantly improve detection accuracy.However,the existing methods usually have complex models and huge computations,which makes them difficult to deploy on resource-constrained devices such as satellites.In addition,due to the shooting height and Angle of the satellite,the scale of the object changes dramatically,so it is inevitable that there are small ships.Despite the rapid development of CNNs,there are still shortcomings in the detection of small target objects.To solve the above challenges,the research content of this paper is as follows:1.This paper proposes a YOLOv5-based lightweight ship detection model called Ship Detection Net,which uses improved convolution units.The improved convolution unit is implemented by applying depthwise separable convolution to replace standard convolution and further using the pointwise group convolution to replace the point convolution in depthwise separable convolution.In addition,the Attention mechanism is incorporated into the convolution unit to ensure detection accuracy.Compared to eleven baseline models and five lightweight models,the proposed Ship Detection Net is more general and effective.2.In order to solve the problem of low detection accuracy of small ships,this paper designs a lightweight small ship detection network based on High-Order Spatial Interactions,called HSI-Small Ship Detection Net.Based on the YOLOv5 framework,this paper selects Ghost Net as its backbone,and adds a tiny target detection branch(P1)on the basis of three target detection branches,so as to enhance the feature extraction of small ships.The HSI-Former based on g~nConv is embedded in the tail of the backbone to enhance the ability of high-order spatial interaction and context modeling.Through the ablation experiment and comparison with other lightweight models,the effectiveness and the superiority of the network are proved.