Mixed pooling and richer attention feature fusion for crack detection

Abstract

• We propose a novel automatic crack detection network. • We design a mixed pooling module. • We adopt channel-wise and spatial attention in different levels. Automatic image crack detection is a critical task for ensuring the safety of various facilities. However, it remains a challenging topic due to the complex background from long and sharp crack topologies. Inspired by recent advances on computer vision applications in deep learning, we propose a novel network architecture with richer feature fusion and attention mechanism and mixed pooling module for crack detection. The proposed network uses the mixed pooling module to replace the conventional spatial pooling. Moreover, we first extract the richer convolutional features to better characterize cracks. Then, we use a spatial attention (SA) in low level feature maps to capture the spatial structure information of cracks. Besides, we use a channel-wise attention (CA) to capture the features of high-level context. Finally, we fuse them together for the final crack prediction. A large crack dataset is used for training and testing. We evaluate our method on a large scale crack dataset, and experimental results on the DeepCrack dataset have demonstrate the effectiveness of the proposed method against state-of-the-art crack detection methods, which achieves Precision( P ) 87.3%, Recall( R ) 88.5%, and F -score over 88%.