Bidirectional Self-Rectifying Networks with Bayesian Modelling for Feature Detection and Keypoint Allocation

• Published in: Proceedings (International Conference on Machine Learning and Cybernetics.) pp. 1 - 6
• Main Authors: Zhu, Qiuchen, Ha, Quang
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.12.2021
• Subjects: Bayes methods; Bayesian Inference; Deep learning; Feature detection; Feature extraction; Hierarchical Convolutional Neural Network; Image Feature Detection; Robustness; Structural panels; Uncertainty
• ISSN: 2160-1348
• DOI: 10.1109/ICMLC54886.2021.9737243

In machine vision, deep learning frameworks are getting more attractive to researchers owing to their accuracy and robustness for feature extraction. However, the uncertainty in data or model has an adversary impact on the prediction and limits the performance of deep learning. To address the problem associated with uncertainty, we propose a bidirectional self-rectifying network with Bayesian modelling (BSNBM) for feature detection. First, a set of branch networks is proposed, wherein the output of previous convolutional blocks is unified and concatenated to the current ones to reduce the visual impairment in the up/down-sampling stage, taking into account the overall information loss. Further, our framework is probabilistically based on Bayesian modelling using prior knowledge. In the Bayesian model, the weight of the learnable layers are converted into distribution functions. Such conversion aims to improve robustness against outliers and there fore alleviate the overfltting issue. The proposed technique is then applied to identify surface cracks of infrastructure such as roads, bridges or pavements. Extensive comparison with existing techniques is conducted on various datasets, subject to a number of evaluation criteria. Experiments on crack images, including those captured by unmanned aerial vehicles inspecting a monorail bridge, demonstrate the merits of the proposed BSNBM architecture over existing techniques for surface defect inspection. Additional tests on extensive applications show the scalability and robustness of this model for various image processing tasks.