Deep learning-based wall crack detection

This study addresses the issues of high weight and complexity in the YOLOv4 network and presents an improved wall crack detection method based on it. The approach involves replacing YOLOv4’s backbone feature extraction network with MobileNetV2 and employing deep separable convolution to reduce model complexity. Additionally, the SENet attention mechanism is integrated to counteract accuracy loss due to lightweighting. The study also includes data set construction and annotation. Experimental results demonstrate that this method significantly reduces network weight, parameters and computational requirements while maintaining high detection accuracy, making it suitable for various wall crack detection tasks. Copyright © 2024 Inderscience Enterprises Ltd.