Quality Detection Of Semiconductor Laser TO56 Based On Machine Vision

With the development of the optical communication technology,optical communication has become the fastest and most stable communication method in the world.As one of the most important basic components in optical communication,the market demand for semiconductor lasers is increased rapidly,and the production quality requirements are became more stringent.TO56 is one of the commonly used semiconductor laser,but its quality inspection is mainly carried on manual sampling inspection,and the production quality is hard to guarantee.Therefore,developing a TO56 automated quality inspection equipment is of great significance to the optical communication component production industry.In recent years,machine vision and deep learning have been successfully applied to quality inspection in various industries.Based on these two methods,a lot of research on the TO56 quality inspection method is carried out in this thesis,an imaging system is designed,and a location-segmentation-defect category pattern recognition three-stage series detection method is proposed in this thesis,and the effectiveness and superiority of each stage algorithm and series detection method are verified through experiments.Finally,the TO56 quality inspection software is developed based on Lab VIEW.The main research work of this paper is organized as follows:(1)Based on the analysis of the internal structural of the TO56,the imaging system and cross rails are designed to realize the automatic imaging of the same batch of TO56.Based on the analysis of the TO56 imaging results and the different defect detection requirements of the chips and the wires,a three-stage series quality inspection process of location-segmentation-defect pattern recognition is proposed.(2)The TO56 chips and wires missing defect detection and location algorithm are studied from the aspects of traditional machine learning and deep learning methods.The contrast experiment results verify that Faster R-CNN is superior to YOLO-v3,SSD,HOG+SVM,and SIFT+SVM algorithm.Aiming at the problem that the wire4 in TO56 image is large aspect ratio detection target and the LD is small area detection target,and the original region proposal parameters of Faster R-CNN cannot effectively adapt to the first stage detection,k-means algorithm is adopted to cluster the ground truth bounding box coordinates of the chips and wires in TO56 images.Through the analysis of the Io U curve based on the cluster centers number and the ground truth,the Faster R-CNN based on optimized region proposal parameters is proposed.The related experiment and comparison results verify the accuracy and superiority of the optimized Faster R-CNN on TO56 chips and wires missing defect detection and location.(3)The traditional image method and deep learning method are used to the wires segmentation of TO56.The experimental results verify that the effect of Deep Lab-v3+is superior to FCN-8s,PSPNet,level set,clustering and threshold segmentation algorithm on each category of wire image segmentation.Aiming at the problem that the solder joint is small and the wire width is narrow target in the TO56 wire image,and the original upsampling process of Deep Lab-v3+ cannot fully regain the detail information of solder joints and wire,the multi-scale idea and SE module are introduced,and the TO56 wire image segmentation algorithm Deep Lab-v3+ based on multi-scale and SE module is proposed.An additional shallow feature whose size is1/2 of the original image is merged in the Deep Lab-v3+ upsampling process to improve the segmentation effect of solder joints and wire,the SE module is adopted to eliminate the redundant features in each channel of the shallow features that contribute little to the improvement of the segmentation effect.The experimental results verify the effectiveness and superiority of the optimized Deep Lab-v3+ algorithm on the TO56 wire images segmentation.(4)Based on the results of TO56 wire image segmentation,the TO56 wire defect category pattern recognition process is designed,and the effectiveness of the locationsegmentation-defect category pattern recognition three-stage series detection method and its related algorithms is overall verified by the test set which is total of 174 TO56 images.The results of segmenting solder joints and wire directly on the original TO56 image without locating and extracting the wire area verify the superiority of the threestage series quality inspection process.Finally,the TO56 quality inspection software was developed based on Lab VIEW,and realize the functions of automatic images collection,detection,and result tracing of TO56.The research work of the thesis has certain theoretical significance and important engineering value for promoting the development of semiconductor laser product quality inspection technology.