Semiconductor Chip Detection And Counting System Based On Deep Learning

Semiconductor chips are widely used in various fields,has become the lifeblood of economic development,national information security,and has a profound impact on modern human life.With the development of target detection technology,deep learning-based semiconductor chip detection and counting has become an important research topic in machine vision.Due to the complicated operation and low precision of traditional manual chip count estimation,it brings great challenge to chip count detection.Therefore,the intelligent research of chip detection and counting is of great significance to promote the development of industry informatization.Based on Anchor-free target detection algorithm,thesis carried out specific research on the semiconductor chip detection and counting problem through the expression of rotating target Angle parameter,feature fusion module,feature alignment module,loss function and lightweight detection model.Specific research contents are as follows:1)An improved FCOS chip detection and counting algorithm is proposed to solve the problem that the semiconductor chip data set is small and the chip target has a dense array and arbitrary direction,which makes it difficult for the horizontal frame detector to accurately detect the chip.This algorithm solves the problem of poor detection accuracy of horizontal frame to chip in any direction by adding Angle prediction parameter,so as to realize precise identification,positioning and counting of chip.The data enhancement strategy was used to expand the number of chip samples to improve the generalization ability of the model.Deformable convolution alignment module and refined regression network are used to solve the problem of misalignment between horizontal receptive field and chip features,and further improve the accuracy of chip Angle recognition.In order to improve the detection accuracy of dense chip,Io U-aware branch prediction and location-related classification scores were introduced into the detection head to sort the detection results.Meanwhile,Varifocal Loss was introduced in the training stage to add the error of Io U prediction into the classification loss function to guide the detection framework optimization.Finally,the improved ATSSR was used to assign appropriate positive sample points to each chip sample,so as to improve the detection accuracy of chips with different sizes.The experimental results show that the improved FCOS target detection network can detect multi-angle chips more accurately,and the recall rate and average accuracy are increased by 3.0% and 4.8% compared with the baseline FCOS model,respectively.2)To deal with the sensitivity of boundary frames to minor deviation of rotation Angle in directed frame target detection,an improved BBAVectors chip detection and counting algorithm is proposed.In order to overcome the problem that the boundary frame is sensitive to the change of rotation Angle,the algorithm avoids adding Angle to chip orientation detection by regression boundary frame edge perception vector.The improved CSPDark Net53 with integrated attention module is introduced to adjust the sensing field of the detection layer and increase the weight of the shallow network feature layer,which effectively enhances the feature extraction capability of the backbone network and the network’s focus on the key areas of chip targets,improves the detection capability of tiny chip targets,and reduces the missed detection of chips in high-density chip scenarios.Fusion expansion convolution is introduced to improve the upsampling structure,expand the scale range of the input of the multi-scale feature fusion network,make the feature map contain the semantic and location information of chips of various sizes,and enhance the detection ability of chips with inconsistent scales.Experimental results show that the recall rate and average accuracy of the improved BBAVectors model are 0.7% and 3.4%higher than the baseline BBAVectors model,respectively.3)Aiming at the problem that the number of parameters of the improved semiconductor chip detection and counting model is too large to be easily deployed to the mobile terminal,a lightweight chip detection and counting algorithm is proposed.By using lightweight backbone feature extraction network Mobile Netv3 and VOVNet,the improved FCOS model was treated as a student model.Then,two knowledge distillation methods,focused global distillation and masked generation distillation,were used to carry out knowledge distillation on the chip detection and counting model,so as to improve the detection performance of the improved FCOS model.The experimental results show that the two improved methods can reduce the number of parameters by 89.56% on the premise of maintaining the accuracy,and realize the lightweight of the detection model,which is more conducive to the deployment of scenarios with limited computing resources and high real-time requirements.Finally,the design and implementation of the chip detection and counting software system is carried out,which is convenient to debug the model,and feedback the detection results,providing efficient and stable detection results,so as to automate the detection.