Abstract
In the task of fabric defect detection, there are problems such as missed detection and false detection caused by defects with large aspect ratios and complex fabric backgrounds. We propose an improved fabric defect detection method based on YOLOv7, which can reduce the amount of network parameters while increasing the detection accuracy. Firstly, a double-branch partial convolution module DBPM is added to the backbone network to reduce the network parameters amount while improving detection accuracy. Secondly, the simple attention mechanism SimAM is introduced in the backbone network to enhance the feature extraction ability of various size and shape defects without introducing additional parameters. Finally, the neck network is reconstructed as a lighter feature fusion network to further reduce the number of network parameters. By testing the datasets, it can be concluded that compared with the original algorithm, the FLOPS of the improved algorithm is reduced by 51.1%, the parameters amount decreases by 36.3%, the mean average precision mAP@0.5 is increased by 5.1%, and the missed detection rate is reduced by 3.6%.