YOLOv5 Algorithm Optimization: Precise Target Detection Amidst Background Interference

This paper solves the challenge that the target is submerged in the background in the border security scenario, and proposes the enhancement of YOLOv5 algorithm. EfficientViT replaces the DarkNet backbone, leveraging Transformer models for improved feature extraction. Modifications to the Spatial Pyramid Pooling-Fast (SPPF) module enhance the algorithm's robustness to scale variations. Introducing SCYLLA-IoU (SIoU) in the loss function improves model convergence speed and accuracy. Validated on a custom border security dataset, the enhancements show an improved average precision mean (mAP) of 48.3%, a 2.2 percentage point increase over the original YOLOv5s model.