Crack-JPU – A crack segmentation method using atrous convolution

Detecting cracks from images using embedded deep learning applications requires efficient and lightweight models in practice. To improve the computational efficiency of models, it is generally aim to reduce the model parameters as much as possible without compromising accuracy. Computational approaches ensure consistency in crack detection across different inspections and operators. Computational methods enable continuous monitoring, including real-time or periodic inspections. The proposed work seeks to leverage the latest deep-learning techniques to get the maximum information out of a minimum number of parameters. The present semantic segmentation-based model - CrackJPU, uses deep hierarchical feature learning convolution networks. Deeply-Supervised Nets (DSN) and JPU (Joint Pyramid Upsampling) modules are also used to supervise the model at multiple inner side-output layers and facilitate retrieval of lower resolution features at decoding layers respectively. To refine the prediction result, the guided filtering method is used. The proposed model has been trained on a standard dataset of annotated crack images. The experimental finding shows that, the model has less than 7 million parameters which are the least compared to recent work without losing performance. Also a mean I/U score of 98.78 and the best F-score is 86.4 is achieved with reduction model parameters. Crack detection is significant in various fields like infrastructure inspection, aerospace industry, Manufacturing Quality Control etc. due to its potential impact on safety, infrastructure integrity, and overall system reliability.