A dual-task algorithm for ship target detection and semantic segmentation based on improved YOLOv5

• Published in: OCEANS 2023 - Limerick pp. 1 - 7
• Main Authors: Fu, Huixuan, Wang, Xiangyu, Peng, Chao, Che, Ziwei, Wang, Yuchao
• Format: Conference Proceeding
• Language: English
• Published: IEEE 05.06.2023
• Subjects: Decoding; dual-task; Feature extraction; Multitasking; Object detection; Sea surface; Semantic segmentation; Semantics; ship semantic segmentation; ship target detection; YOLOv5
• DOI: 10.1109/OCEANSLimerick52467.2023.10244376

Marine target detection technology plays an important role in sea surface monitoring, sea rescue and other fields. Aiming at the problem of low detection accuracy rate of ship targets, this paper proposed a dual-task network model which integrates target detection and semantic segmentation for ship objects. The semantic segmentation branch composed of Encoder and Decoder is added on the basis of YOLOv5 network to obtain multi-scale context information. Encoder and Decoder are connected through ASPP module, finally the low-level features and high-level features are further fused to generate semantic segmentation results. CE Loss is used as the semantic segmentation loss function to perform gradient optimization. Finally, it is superimposed with the target detection results to generate an image with both the anchor frame information of the target position and the contour information of the target boundary as the final result. The two tasks share the backbone extraction network and feature fusion network of YOLOv5. Different tasks have different positions of local minima in the shared layer. The interaction of unrelated parts between multiple tasks helps to escape from local minima. The detection accuracy is improved through the features of information sharing and mutual complement of multi-task models, and the semantic segmentation function is implemented. The experimental results show that, on the self-built dataset, the mAP of target detection is improved by 4.3 percentage points compared with YOLOv5, and the MIoU of semantic segmentation can reach 93.6%.