YOLO-SO: Small Object Detection Research Based on YOLOv5s for Crack Detection

Abstract

Addressing issues such as slow speed, low accuracy, and insufficient generalization performance of traditional small target detection algorithms, this paper proposes a novel real-time detection method for small targets named YOLO-SO based on the YOLOv5s deep learning object detection algorithm, and building cracks are taken as the research target. This method optimizes and improves the YOLOv5s object detection neural network. Firstly, CBM (Conv + BN + Mish) depth separable convolution modules are introduced into the backbone network layer, and lightweight CA (Coordinate Attention) is added to the output feature map of the backbone network to focus more on crack features, thus enhancing detection performance. Secondly, the dense connection concept is introduced, replacing the feature fusion network with the PADNet network to reuse feature information. Finally, the Complete Intersection over Union (CIoU) is introduced as the target localization loss function. Experimental evaluations are conducted on a crack dataset using Mosaic data augmentation, and comparisons are made with various existing object detection neural networks. The experimental results demonstrate that the improved model, compared to the original model, reduces parameter volume by 43.28%, reduces computational load by 47.47%, and improves detection accuracy by 2.18%, validating the superiority of the proposed algorithm in this paper.