Abstract: Object detection algorithms based on deep learning have been fully promoted and applied in the field of industrial defect detection,but few algorithms are suitable for single background in industrial detection scenarios.This paper takes the industrial detection scene with a large number of simple repeated backgrounds as the starting point,and makes the following improvements to the RetinaNet algorithm:1)introduce the difficult negative sample mining strategy to reduce the impact of a large number of simple repeated negative samples on the model fitting positive samples;2)an adaptive ignoring sample selection strategy is designed to avoid mixing samples with high intersection ratios of positive samples into negative samples to confuse model training;3)the classification sub-network of RetinaNet is simplified,and the risk of overfitting after model improvement is reduced.Compared with the RetinaNet algorithm,the improved method improves the recall rate by 14.1% and 1.8%,and the precision rate by 3.6% and 0.4% respectively on the public PCB missing hole dataset and the self-built LED bubble dataset,indicating that the improved method can significantly improve the level of object detection in a single background.

Key words: Deep learning, RetinaNet, Adaptive sampling, Single background