| With the steady progress of industrial intelligence and the increasing scale of intelligent storage,inventory work intelligence has become an urgent need.Almost all current work is inseparable from the PCB,the supply of its raw materials double-sided copper clad laminate is very high.The traditional inventory method is difficult to meet the inventory precision and speed of double-sided copper clad laminate,thus seriously affecting the production and processing efficiency.In practical applications,existing image recognition algorithms still face many challenges,such as the lack of available double sided copper clad laminate data set resources,image recognition algorithm model precision is poor,and the lack of timeliness and other problems.In this thesis,we conduct an in-depth study to address these issues and propose an an image recognition-based algorithm for double sided copper clad laminate stocktaking in PCB warehouses to improve the performance and robustness of the model.The following work and contributions are accomplished during the research.(1)To address the current challenge of the scarcity of double-sided copper-clad PCB warehouse datasets,a dataset of 1483 images containing more than 10,000 double-sided copper-clad targets is independently produced.This dataset contains different situations in various practical application scenarios,such as different lighting and background conditions,to enhance the robustness of the algorithm.(2)To address the problem of insufficient detection precision for double-sided copper-clad panels in the current YOLOv5 algorithm,an improved method based on the attention mechanism is proposed.In the backbone part of the algorithm model where the C3 module is improved using the SE attention mechanism approach.The improvement can make the model pay more attention to the detection target and have a stronger ability to distinguish the target recognition,so as to improve the recognition precision of the model,and the effectiveness of the improved way is verified through experiments.(3)In order to reduce the computation of the algorithm model in the detection process,an improvement method based on a lightweight network model is proposed.The backbone network of the algorithm model is optimized with multiple lightweight models and comparison experiments are made.Next,the activation function is improved to Hard Swish to reduce the amount of redundant operations.It is demonstrated that the improved strategy based on Ghost Net lightweight model can effectively reduce the computational power of the algorithm,greatly reduce the size of the number of parameters and significantly improve the model detection speed.(4)Proposed a fusion improvement YOLOv5 method based on the improved Ghost lightweight network and SE attention mechanism.At the same time,a human-computer interface based on the Flask micro-framework is designed and developed to visualize the operation of the double-sided copper clad inventory.The experimental results prove that the fused improved YOLOv5 algorithm model has a very good performance effect in terms of precision as well as speed. |
