A Supervised Contrast Focus Neural Network for LED packaging defects classification

Unbalanced classification of multi-defects in LED packaging is a challenging and meaningful task. The imbalance problem between positive and negative samples is caused by the high collection cost and the limit on the number of defective samples in industrial production. Defects are diverse and complex in form, and different types of defects are interleaved. Supervised Contrast Focus Neural Network (SCFN) with two sub-network is proposed to solve these problems. The two sub-networks of SCFN are responsible for identifying and classifying defects, respectively. Focus Net (FN) sub-network provides zero-defect prior knowledge to the classifier sub-network and integrates it into the process of decision-making to reduce the class imbalance ratio and resist the effect of simple positive samples. In addition, a collaborative loss is constructed to coordinate supervised contrast and cross-entropy to alienate the between-class distances of close samples. Experiments show that the proposed method achieves 93.47% (defect classification) in the extremely unbalanced defect classification tasks. Ablation experiments demonstrate the effectiveness of the proposed approach in combating imbalances.