Research On Defect Detection Algorithm Of Chip After WB Process Based On Deep Learning

In the era of artificial intelligence,integrated circuit chips are indispensable for applications in various fields such as communications equipment,consumer electronics,aerospace,and industrial manufacturing.Therefore,the demand for chips is increasing,and the quality requirements are gradually increasing.Integrated circuit testing is a key link in the whole process of the chip industry chain through design,manufacturing,packaging,and application,and it is related to the performance,quality and life of the finished chip.Due to the small size of the chip,high precision,and large demand,the accuracy and efficiency of chip testing cannot be ignored and technically difficult in chip production.In the process of chip production,manufacturing and packaging,wire bonding(WB)is one of the most critical processes,which directly determines the availability and stability of the finished product.After the chip WB process,the most common types of defects are: missing wire,broken wire,insufficient bond area,etc.Therefore,this article will detect the above three defects of a power amplifier chip.First,the traditional vision method is used for detection,and the realization of the chip vision system based on this method is demonstrated,and the limitation of the traditional method to solve the target detection problem is analyzed.Then,according to the related research and applications of deep learning that has been continuously developed in recent years,this paper proposes a deep learning-based chip WB defect detection algorithm after the process.The specific research content is as follows:Using the YOLO(You Only Look Once)algorithm in target detection combined with the idea of optimization of anchor frame calculation and migration training,an experimental analysis was carried out on the chip defect data set in this paper,and the results proved that the method can be applied to the chip detection task.At the same time,an online target detection system was developed based on the overall detection process of this experiment.Aiming at some optimizable directions of the adopted YOLO algorithm on the sample data set used in this experiment,the introduction of an attention mechanism is proposed to embed the sc SE(Concurrent Spatial and Channel Squeeze & Excitation)module into the backbone of the YOLO network.The fusion area allows the mean Average Precision(m AP)to increase by 1.6% while keeping the model volume without significant increase.For the possible needs of real-time detection and deployment of embedded devices in practical applications,the Center Net target detection algorithm using the idea of anchorless frame is tested on the dataset of this paper,and improvements are made based on this algorithm by introducing a depth-separable convolution in Res Net-50 to replace the original feature extraction network Res Net-18,which reduces the number of model parameters and improves the detection The number of model parameters is reduced and the detection speed is improved.