A YOLO-CDA X-ray Image Detection Algorithm
Yongjian Li, Huasheng Zhu, Mingzhi He, Shuyin Tang, Zhanxin Sun · 2023
It is difficult for the YOLO algorithm to accurately detect and locate small target contrabands that are similar to the background and occlude each other in security detection applications. In response to this problem, this paper proposes a YOLO-CDA algorithm, which mainly improves the upsampling module, embeds channel convolution self-attention and spatial convolution self-attention based on the CARAFE structure, and strengthens the new upsampling operator to capture the spatial and channel dependencies between long-distance features, effectively establishes long-distance contextual connections, and ensures the integrity of semantic information for contraband imaging and background distinction and upsampling. At the same time, a new online data augmentation method is introduced in the pre-processing to increase the linear combination of images to alleviate the over fitting phenomenon, and also increase the diversity of input training images. The results show that the proposed YOLO-CDA algorithm mAP0.5:0.95 reaches 76.9%, which is 4.7% higher than the accuracy of the original model algorithm; FPS reaches 63, which meets the real-time detection requirements.