Abstract Regarding difficulty and low accuracy in closed-set recognition (CSR) of the ceramic welding pad defects, a dense visual representation method of multi-scale feature tensor for images was proposed. Based on this method, the fine granularity of normal samples can be finely captured and relevant calculations for subsequent probability transformation can be performed. By designing a probability transformation module of multi-scale feature tensor for images and cascading multiple modules increasing parameters of the model, a probability transformation model of multi-scale feature tensor for image was established. It was determined whether an image falls within or outside the distribution, thus achieving high-accuracy open-set recognition (OSR) of abnormal image samples. This OSR method was applied and implemented for the detection of ceramic pad defect samples. Finally, performance comparison experiment and anti-interference experiment were conducted with the OSR method based on Gaussian mixture model (GMM). The experimental results show that the method based on GMM has an Area Under Curve (AUC) of 0.99, while the method proposed in this paper has an AUC of 1, representing an improvement of 1.01%. Moreover, the proposed method exhibits a more pronounced and superior separation effect, achieves very accurate defect localization, and demonstrates excellent noise resistance.
Wan et al. (Sat,) studied this question.