Multi-Scale Feature Similarity and Object Detection for Small Printing Defects Detection

Abstract

It is always a challenging task in the industry to detect the small printing defects under complex background. To address this problem, a defect detection algorithm based on multi-scale feature similarity evaluation and small object defect detection is proposed. Firstly, we use a Siamese neural network to extract the multi-scale features of reference image and detection image. Multi-scale features are used to characterize the background information and defect object information of printing image. Secondly, to segment defect object features from the complex background features, we analyze the differences between reference image features and detection image features by calculating the similarity heat map. Further, we can get a series of candidate area with possible defects. Finally, we use a decoupled head to decode the features in the candidate region. Experimental results show that the proposed algorithm can accurately detect small printing defects in complex background and reduce the false positive rate of the detection system.

Keywords

• Defect detection
• multi-scale feature similarity evaluation
• Siamese network