Deformer-FPN: FPN network based on deformable convolution and lightweight upsampling operator for ship detection in SAR image

This paper proposes a Deformer-FPN target detection network based on the minimum model size yolov5-v7 (yolov5s) for detecting ships of different scales in SAR images. Due to the varied size of the ship and the different resolution of the image, ships exist in a multi-scale space in the SAR image. To address this issue, we enhance the feature extraction of the target in complex backgrounds and improve the adaptability to ship size. Firstly, we incorporate the latest deformable convolutional DCNv3 into the input part of the existing Feature Pyramid Network (FPN, a neural network architecture). This integration enhances feature extraction in complex backgrounds and improves adaptability to ship size. Additionally, we introduce an additional small-size feature reinforcement branch to further strengthen the extraction of small-size targets. Secondly, we introduce a lightweight general-purpose upsampling operator (CARAFE) to capture global spatial semantic information of ship features, addressing the limitation of nearest neighbor upsampling that only considers adjacent pixels. Experimental results on the SAR ship detection open source dataset (SSD) demonstrate the effectiveness of the proposed method. The Yolov5s+Deformer-FPN detection algorithm achieves an average accuracy of 91.89% and a recall rate of 92.47%. The experimental code reference is as follows: https://github.com/Maple1024/Deformer-FPN.git