Structured Feature Representation and Unbalanced Category Optimization for CrackDetection
Jiaqiu Guan, Yuting Hang, Jie Fang, Zhijie Zhu · 2024
Even though crack detection methods based on deep neural network have improved the inference performance and efficiency significantly, it is still limited by complex background interference and unbalanced category distribution, it hence can not satisfy the practical requirements. In this case, we propose a structured feature representation and unbalanced category optimization (SFRUCO) based crack detection framework, which addresses the abovementioned issues from two aspects of network architecture and majorization strategy. Specifically, we present a multi-branch feature representation module as the backbone network to encode the dependencies among samples at different levels. At the same time, the module combines a convolutional block attention module (CBAM) to enhance the robustness and discrimination of the final feature. Besides, we present a two-stage marching inference mechanism to refine the detection result in a coarse-to-fine fashion to decrease the detection difficulty. In addition, we present a dynamical cost function to balance the sensitivities of the model for different categories adaptively in the training phase and avoid the risk of model collapse. Finally, the experiments validate the proposed SFRUCO is of effectiveness and superiority.