Rail Flaw Detection Method Based on Yoloac

Railway transportation is crucial in the transportation system, but rails can sustain various damages during long - term use, endangering train safety. Ultrasonic flaw detection, a common method to detect rail damage by analyzing returned ultrasonic signals, faces issues such as small and indistinguishable ultrasonic signal images for some damages due to their small size and ultrasonic detection sensitivity settings, resulting in low detection accuracy with existing deep - learning methods. To address this, this paper proposes an improved small - object detection network, the YOLOAC model, based on YOLOv9. It combines attention mechanism and contextual learning, uses PGI from an auxiliary network to update weights, abandons the original YOLOv9 convolutional neural network in the backbone for the TransNeXt network with the Aggregated Pixel-focused Attention mechanism, adds the Container context aggregation network for context extraction, and the CBAM attention mechanism in the Auxiliary Reversible Branch for more reliable gradient information. Experimental results show that YOLOAC outperforms existing models in rail flaw detection, with its <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$m A P_{50} 0.009$</tex> higher than YOLOv9 - <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{C}$</tex> and 0.013 higher than YOLOv10 - M, and ablation studies confirm the effectiveness of the improvements.