| With the rapid development of the global information technology industry,LED chips have been widely used in many fields such as indicators,signs,lighting and biological detection,etc.Whether the chip package meets the quality requirements directly affects the appearance and usage of the finished product.As deep learning methods can effectively extract to image detail information,they have been widely used in the field of target detection to provide new ideas for defect detection in the chip surface packaging process,but the existing deep learning detection methods cannot achieve the balance of speed and accuracy.Therefore,this paper combines the needs of the production process of enterprises to carry out the research of chip surface defect detection method,which is not only of theoretical significance but also of more application value.The research work accomplished in this paper are as follows.(1)In order to solve the poor accuracy of the SSD algorithm when detecting chip surface defects,and to address the characteristics of the LED dispensing dataset such as small number of samples,small scale of defects,and inconspicuous features,an improved SSD algorithm for LED chip defect detection,AT-SSD,is proposed.Specific improvements are as follows: 1)the VGG feature extraction network is replaced with a deep residual network,Res Net,and a series of feature layer structure;2)in order to focus on effective features in both spatial and channel dimensions,the feature attention module CBAM is added to select appropriate features;3)to solve the positive and negative sample imbalance problem,the Focal Loss loss function is introduced;4)by analyzing the characteristics of different datasets with different aspect ratios,optimization in training strategy is done.Through comparison and ablation experiments,m AP improves 4.3% compared with the original SSD,which verifies the effectiveness of the algorithm in performing chip surface defect detection.(2)In order to fuse the contextual semantic relationships between different feature layers,a multi-scale feature fusion-based LED chip defect detection algorithm F-AT-SSD is proposed.This algorithm proposes a lightweight feature fusion module based on the AT-SSD architecture,which further improves the accuracy of LED chip surface defect detection accuracy without excessive loss of speed,and enhances the robustness of the network.The experiments obtained 79.7% m AP and 78.7% m AP on the LED dispensing dataset and PASCAL VOC2007 test set,respectively,which verified the effectiveness of the improved method. |
