A multiple position-based bi-branch model for structural defect inspection

Reliable industrial defect inspection is one of the main challenges in manufacturing scenarios, especially for the inspection of structural defects. However, the cost of missing a defect is much higher than the cost of misclassifying a qualified sample, which is seldom emphasized in previous work. Thus, the purpose of our work has two folds: reduce the omission rate of defective samples; classify industrial samples correctly. To that end, in this paper, we first define a position tag for each sample, where samples with the same position tag describe the same product information. We also design the multi-position weighted-resampling (MPWR) method for extracting paired data with identical tags. Then, in order to fully learn from the paired data, we propose a multiple position-based bi-branch (MPB3) neural network architecture to perform similarity measurements and multi-classifications simultaneously. Experimental results demonstrate the effectiveness of our method and generalization capacity to data from unknown tags by comparing with other methods. For example, the proposed method achieves 2.77 $$\%$$ % /1.00 $$\%$$ % omission rates and 96.81 $$\%$$ % /99.03 $$\%$$ % weighted F-Scores on the SMT defect dataset and the motor brush holder dataset, respectively. In addition, the average running time of the method only needs 9.6 ms, which meets requirements of cycle time in manufacturing industries. In conclusion, the omission rate of defective samples can be reduced effectively by the position-based method that consists of MPWR method and MPB3 structure, which greatly improves productivity in real production lines.