Surface Crack-specific Knowledge Localization in Pretrained Convolutional Neural Networks

This paper presents analysis of distribution of knowledge in pretrained convolutional neural networks useful for surface crack detection. Surface crack detection is in focus of many researchers as its application to building inspection and structural health estimation, crucial for safe operation of civil infrastructure. Previous works indicate high performance obtained by using transfer learning on popular pretrained convolutional neural networks architectures such as AlexNet, GoogLeNet, ResNet, VGG. Goal of this paper is to further investigate where application-specific knowledge is contained in pretrained networks. Pretrained networks were systematically reduced and partially retrained. Performance analysis on publicly available “Concrete Crack Images for Classification” dataset indicate that in majority of cases useful knowledge is contained in first layers of networks. Reduction of memory footprint obtained without loss of accuracy is at least an order of magnitude. Inference time reduction is consequently achieved together with memory footprint reduction.