A Lightweight Network for Crack Detection With Split Exchange Convolution and Multi-Scale Features Fusion
Qiang Zhou, Zhong Qu, Fang-Rong Ju
Abstract
Crack detection mostly benefits from the rapid development of Convolutional Neural Networks (CNNs). However, the improvement of crack detection performance comes from the deeper and wider network structure, which requires heavier computation and storage overhead. This prevents crack detection methods from being deployed on practical platforms, especially mobile devices. To tackle this problem, we propose a novel Split Exchange Convolution (SEConv) modules, which splits the feature maps into high resolution and low resolution parts and then filters out the redundant information of each part. SEConv exchanges the feature information between the two modules to make the feature efficient reuse. Besides, we design a Multi-Scale Feature Exchange (MSFE) module to promote the cross stage features fusion. Benefiting from the SEConv and the MSFE modules, we build an extremely lightweight crack detection model with only 1.3 M parameters and 8 G <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">FLOPs</i> while achieving comparable performance. Extensive experimental results on the crack detection benchmark show that our method consistently outperforms other state-of-the-art methods in the evaluation metrics of <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">F1-score</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MIoU</i> .