A lightweight convolutional neural network for detecting road cracks.

Autor: Xinghua, Ren, Shaolin, Hu, Yandong, Hou, Ye, Ke, Zhengquan, Chen, Zhengbo, Wu
Zdroj: Signal, Image & Video Processing; Sep2024, Vol. 18 Issue 10, p6729-6743, 15p
Abstrakt: One of the key procedures in road maintenance is crack detection. Due to limitation in the memory and computing resources of detection devices, existing general detection models, such as YOLOv7-tiny, are considered among the most advanced technologies in the field of road crack detection. However, they still require optimization when applied to real-world detection scenarios. In this article, an efficient and lightweight network model YOLO-DC for road crack detection is designed. In this model, an Efficient Coordinate Attention cross-stage backbone Network composed of the cross-stage learning module, the Coordinate Attention mechanism and the Partial Convolution is proposed. Then, multi-scale features are used to enhance the quality of feature extraction and the representation for solving the problem of small crack detection. Then, a novel Cross-stage Bidirectional Feature Pyramid Network is designed based on multi-scale features to improve the detection accuracy with smaller computational load. Finally, the knowledge distillation is performed to further improve the detection accuracy without increasing the computational load. In addition, the effectiveness of the designed lightweight detection network is evaluated by comparing the experimental results on the self-made data set. Compared to the State-Of-The-Art model, our proposed model is 0.4 % lower in precision, but 69 % less in the parameters and calculation amount. At the price of slightly lower the detection accuracy, the model parameters and complexity are greatly reduced. This is conducive to the application of the model in road crack detection. [ABSTRACT FROM AUTHOR]
Databáze: Complementary Index
Externí odkaz: