Adaptive convolution confidence sieve learning for surface defect detection under process uncertainty

Abstract

Surface defect detection plays an important role in the quality management of industrial manufacturing processes. Existing detection methods are developed based on clean and uncontaminated training data, less considering the process uncertainty. This paper proposes an adaptive convolution confidence sieve learning method for surface defect detection under process uncertainty. First, an edge-separated wavelet transform method combines the spatial and transform domain methods to estimate image noise. After that, an adaptive convolution model integrates parallel multiple convolution kernels, which can cover the defect feature variation. Then, a confidence sieve learning scheme is developed to filter the label noise, and the adaptive convolution is fine-tuned on the high-confidence data to achieve better detection accuracy. Finally, experimental studies on three industrial datasets have verified that the edge-separated wavelet transform method can accurately predict image noise. The adaptive convolution has better representation capability, and the confidence sieve learning improves model robustness.