A Decoupled Head and Multiscale Coordinate Convolution Detection Method for Ship Targets in Optical Remote Sensing Images

• Published in: IEEE access 2024, Vol.12, p.59831-59841
• Main Authors: Ma, Jin, Fu, Dongyang, Wang, Difeng, Li, Yongze
• Format: Article
• Language: English
• Subjects: Algorithms; Convolution; coordinate embedding; Datasets; decoupled head; Feature extraction; Feature maps; Fisheries; Machine learning; Marine environment; Marine vehicles; multi-scale; Object detection; Object recognition; Optical imaging; Optical sensors; Performance measurement; Remote sensing; Target detection; YOLO
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3394052

Intelligent detection and recognition of ship targets is the basis of naval battlefield situational assessment and threat estimation. Its core task is to determine whether there is a ship target in the image and to detect, identify, and locate the ship target, which also has broad application prospects in fisheries management, maritime rescue, maritime traffic management, and marine environment monitoring. To this end, a multi-scale ship target detection improved algorithm for remote sensing images is proposed based on the YOLOv5 model, called CPE-YOLO. Firstly, a multi-scale attention fusion module based on coordinate position is proposed to solve the problem of missing spatial perception ability for ship detection in optical remote sensing images, to process the multi-scale coordinate information of feature maps, and to effectively establish long-range channel dependency among multi-scale channel attention. Secondly, a more refined feature pyramid network is constructed to effectively mitigate the impact of target scale variation due to remote sensing images on model performance, and to provide richer feature information for feature fusion layer regression localization. Ultimately, the decoupled head is integrated into the YOLO Head to disentangle the classifier and regressor, facilitating accelerated network convergence and concurrent enhancement of network detection accuracy. Finally, thorough tests on the HRSC2016, ASDC and SIGE datasets are presented to support our methodology. The results show that the performance of our proposed methods is better than other existing methods, with an mAP of 94.0%, 76.5%, and 98.7% on the HRSC2016, ASDC, and SIGE datasets. Moreover, comparative assessments against alternative deep learning methodologies affirm the sustained superiority of the enhanced method across overall performance metrics.