부산대학교 도서관

Traditional polarity detection of welded chip mainly adopts Automated Optical Inspection (AOI) detection, and the chip with opposite polarity is manually rechecked. AOI detection speed is fast, but the false detection rate and missed detection rate are high, and the amount of manual rechecked data is very large. To this end, this paper builds a deep neural network integrating attention mechanism and feature fusion module. Based on the data set provided by enterprise, an active learning algorithm based on the query strategy of truth degree (the confidence of the unlabeled data set detection category and the positioning judgment correct) and falsity degree (the confidence of the unlabeled data set detection category or the positioning judgment wrong) is designed to constantly update the training set, train and optimize the deep neural network. A post-processing mechanism considering the position relationship between the polarity mark of chip and the polarity mark of PCB board is designed to achieve the high-quality detection of polarity defects of chip welding and greatly reduce the workload of manual reinspection. The model was tested on the test set, and the recall rate was 100 % , the accuracy rate was 90.7 % , and the processing time of single image was 5ms.