DENS-YOLOv6: a small object detection model for garbage detection on water surface

• Published in: Multimedia tools and applications Vol. 83; no. 18; pp. 55751 - 55771
• Main Authors: Li, Ning, Wang, Mingliang, Yang, Gaochao, Li, Bo, Yuan, Baohua, Xu, Shoukun
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.05.2024; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Datasets; Deep learning; Garbage; Interference; Modules; Multimedia; Multimedia Information Systems; Noise reduction; Object recognition; Special Purpose and Application-Based Systems; Telematics; Track 6: Computer Vision for Multimedia Applications; Water waves; Small object detection; Adaptive noise suppression; YOLOv6; Water surface garbage detection; Detail information enhancement
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-023-17679-7

The study of garbage detection on water surface is of great significance for the development of water surface garbage monitoring and automated water surface garbage salvage. However, in water surface garbage scenes, the proportion of water background is relatively large, while the proportion of detection objects is relatively small. Moreover, the objects are easily affected by noise interference such as lighting, water waves, and reflections, which makes it difficult to extract object features and affects detection accuracy. In this paper, we propose a Detail Enhancement Noise Suppression YOLOv6 (DENS-YOLOv6) detection algorithm based on YOLOv6. Firstly, to better capture the detailed feature information of small objects, we design a Detail Information Enhancement Module (DIEM) based on atrous convolution. Secondly, to suppress noise interference on small objects, we develop an Adaptive Noise Suppression Module (ANSM). Finally, in order to improve the stability and convergence speed of the model training, we employ a regression loss function based on the Normalized Wasserstein Distance(NWD) metric. Experiments were conducted on the Flow+ dataset with a large number of small objects and the publicly available Pascal VOC2007 dataset. The mAP S indicators reached 40.6% and 11.4%, respectively. Compared with other models, DENS-YOLOv6 achieved the highest small object detection accuracy