@Article{csse.2024.056736,
AUTHOR = {Conghao Niu, Dezhi Han, Bing Han, Zhongdai Wu},
TITLE = {SAR-LtYOLOv8: A Lightweight YOLOv8 Model for Small Object Detection in SAR Ship Images},
JOURNAL = {Computer Systems Science and Engineering},
VOLUME = {48},
YEAR = {2024},
NUMBER = {6},
PAGES = {1723--1748},
URL = {http://www.techscience.com/csse/v48n6/58696},
ISSN = {},
ABSTRACT = {The high coverage and all-weather capabilities of Synthetic Aperture Radar (SAR) image ship detection make it a widely accepted method for maritime ship positioning and identification. However, SAR ship detection faces challenges such as indistinct ship contours, low resolution, multi-scale features, noise, and complex background interference. This paper proposes a lightweight YOLOv8 model for small object detection in SAR ship images, incorporating key structures to enhance performance. The YOLOv8 backbone is replaced by the Slim Backbone (SB), and the Delete Medium-sized Detection Head (DMDH) structure is eliminated to concentrate on shallow features. Dynamically adjusting the convolution kernel weights of the Omni-Dimensional Dynamic Convolution (ODConv) module can result in a reduction in computation and enhanced accuracy. Adjusting the model’s receptive field is done by the Large Selective Kernel Network (LSKNet) module, which captures shallow features. Additionally, a Multi-scale Spatial-Channel Attention (MSCA) module addresses multi-scale ship feature differences, enhancing feature fusion and local region focus. Experimental results on the HRSID and SSDD datasets demonstrate the model’s effectiveness, with a 67.8% reduction in parameters, a 3.4% improvement in AP (average precision) @0.5, and a 5.4% improvement in AP@0.5:0.95 on the HRSID dataset, and a 0.5% improvement in AP@0.5 and 1.7% in AP@0.5:0.95 on the SSDD dataset, surpassing other state-of-the-art methods.},
DOI = {10.32604/csse.2024.056736}
}