Fusion of Adaptive Convolutional Neural Networks for SAR Images in Ship Target Detection Methods

‍ ‍Synthetic aperture radar （SAR） ship target detection is an important application of microwave remote sensing； however， as the sea surface， especially offshore background is becoming more complex and the sizes of the ships are different， degradation of the performance of the detection algorithm， leakage， incorrect detection， and other problems crucially affect the detection of the target. With the aim of solving the abovementioned problems， this paper proposes a fusion adaptive convolutional neural network ship detection method for SAR images in complex scenes. First， to improve the feature extraction ability of the backbone network， an adaptive Omni-Dimensional Dynamic and Efficient Layer Attention Network （OD-ELAN） module for feature extraction is designed. This module is a dynamic convolution fused in an efficient layer attention network （ELAN）， which enables the adaptive OD-ELAN module to dynamically adjust the weights of the convolution kernel with different input features. Second， to improve the network’s ability to localize the target， the explicit visual center （EVC） mechanism was used in the head network to enhance the network’s ability to perceive different regions spatially， enabling it to adapt more flexibly to features of different sizes， scales， and locations. The head network structure was then designed to reduce the computational effort of the network， while maintaining the accuracy of network detection. Finally， the Smoothde Intersection over Union（SIOU） loss function was used to measure the degree of overlap between the predicted bounding box and the real bounding box to improve the network’s ability to localize the detected target. To verify the effectiveness of adaptive convolutional neural networks， ablation experiments and comparison experiments were conducted using the SAR Ship Detection Dataset（SSDD） and the High Resolution SAR Images Dataset（HRSID）. The results of the ablation experiments showed that the detection effect of the network proposed in this paper was improved after each improvement. The improved network enhanced the average precision of SSDD dataset ship target detection by 7.81%， accuracy by 4.95%， and recall by 11.8%. For the HRSID dataset， the average precision of ship target detection was improved by 10.6%， accuracy by 10.03%， and recall by 11.84%. The experimental results showed that the improved algorithm significantly reduced false and missed detections when detecting ship targets in small targets and complex backgrounds.