Multi-level feature enhancement network for object detection in sonar images

The unstable geometric features affect the accuracy of object detection in sonar images. We thus propose a novel multi-level feature enhancement network to enhance useful features for object detection in sonar images. We first introduce a deformable convolution to model variations in geometric features. In addition, spatial and channel attention modules are utilized to aggregate rich semantic information from features, improving the quality of feature extraction. We further use an adaptive multi-scale feature fusion module for feature weighting so as to enhance fine-grained features and minimize information loss during feature fusion. Then, the cascaded detection module corrects the prediction results of the previous detector with a low Intersection-over-Union (IoU) threshold, where each detector employs adaptive feature enhancement blocks to enhance region proposal features and thus improve detection performance. Experimental results on two real-world sonar image datasets show that our proposed model performs better than several mainstream object detection methods by achieving 2% to 19.4% higher accuracy rates.