Multi-scale attention networks for pavement defect detection

• Published in: IEEE transactions on instrumentation and measurement p. 1
• Main Authors: Chen, Junde, Wen, Yuxin, Nanehkaran, Yaser Ahangari, Zhang, Defu, Zeb, Adan
• Format: Journal Article
• Language: English
• Published: IEEE 22.07.2023
• Subjects: attention module; Convolution; Convolutional neural networks; deep neural network; Feature extraction; image identification; Kernel; multi-scale convolution; Pavement defect detection; Roads; Task analysis; Training
• ISSN: 0018-9456
• DOI: 10.1109/TIM.2023.3298391

Pavement defects such as cracks, net cracks, and pit slots can cause potential traffic safety problems. The timely detection and identification play a key role in reducing the harm of various pavement defects. Particularly, the recent development in deep learning-based CNNs has shown competitive performance in image detection and classification. To detect pavement defects automatically and improve effects, a multi-scale mobile attention-based network, which we termed MANet, is proposed to perform the detection of pavement defects. The architecture of the encoder-decoder is used in MANet, where the encoder adopts the MobileNet as the backbone network to extract pavement defect features. Instead of the original 3×3 convolution, the multi-scale convolution kernels are utilized in depth-wise separable convolution layers of the network. Further, the hybrid attention mechanism is separately incorporated into the encoder and decoder modules to infer the significance of spatial points and inter-channel relationship features for the input intermediate feature maps. The proposed approach achieves state-of-the-art performance on two publicly-available benchmark datasets, i.e., the Crack500 (500 crack images with 2,000×1,500 pixels) and CFD (118 crack images with 480×320 pixels) datasets. The mean intersection over union ( MIoU ) of the proposed approach on these two datasets reaches 0.7219 and 0.7788, respectively. Ablation experiments show that the multi-scale convolution and hybrid attention modules can effectively help the model extract high-level feature representations and generate more accurate pavement crack segmentation results. We further test the model on locally collected pavement crack images (131 images with 1024×768 pixels) and it achieves a satisfactory result. The proposed approach realizes the MIoU of 0.6514 on the local dataset and outperforms other compared baseline methods. Experimental findings demonstrate the validity and feasibility of the proposed approach and it provides a viable solution for pavement crack detection in practical application scenarios. Our code is available at https://github.com/xtu502/pavement-defects.