Research On Multi-scene Electronic Component Detection Technology Based On Deep Learning

In recent years,artificial intelligence and intelligent manufacturing have witnessed a vigorous development.The key machine vision technology is widely used in various production lines,which can not only improve efficiency and reduce costs,but also reduce labor intensity of workers and improve the level of enterprise intelligence.Currently,the assembly of electronic components mostly relies on manual insertion and assembly,which is not only time-consuming and laborintensive,but also costly and inefficient.Therefore,based on machine vision technology to achieve automatic assembly of electronic components has become the key.The electronic components are small in size and weak in features,and the same electronic component has a large difference in features due to the change of pose in different assembly scene.Building separate detection models for each assembly scene will lead to frequent switching of models to adapt to different scenes during the assembly process,which greatly affects the assembly efficiency.The detection of electronic components in multi-scene is a key step to realize the assembly of electronic components,but there is a lack of reliable solutions.In view of the above problems,this thesis carries out the research of multi-scene electronic component detection methods,the main research content and related conclusions are as follows.(1)Research on high-precision detection methods for electronic components in multi-scene.Firstly,a multi-scene electronic component detection dataset is constructed,which contains three scenes before,during and after assembly.Then the object detection methods of the backbone feature extraction network,feature fusion network,prediction network,and auxiliary training methods are studied in depth respectively,and the method for high-precision detection of electronic components in multi-scene is proposed.Meanwhile,considering that the performance of the model trained by the Io U threshold-based positive and negative sample matching method is affected by the set hyperparameters(Io U threshold),an adaptive K-Means-based positive and negative sample matching method is proposed,which does not need to set hyperparameters,but optimizes and adjusts its own method simultaneously during the network training process.Finally,comparison experiments are conducted on the public dataset Pascal VOC and the constructed multi-scene electronic component dataset,and the results show that the proposed method has outstanding performance,achieving the highest m AP(83.44% and 98.83%),lower computational volume(44.26GMac)and smaller number of parameters(29.3M),with a multi-scene electronic component detection task of higher accuracy.(2)Research on lightweight electronic component detection method based on knowledge distillation.In order to further achieve rapid and accurate detection of electronic components in multi-scene,a lightweight electronic component detection model is first constructed.Then,in order to compensate for the decrease in accuracy caused by the lightweight model,the constructed high-precision detection model is used as the teacher network and the constructed lightweight model is used as the student network,and then the student network accuracy is improved by using the knowledge distillation method.Considering the problem of expression differences between the teacher network and the student network,and in order to learn the rich class-relevant features of the teacher network while also allowing differential learning of class-irrelevant features,a knowledge distillation method based on the combination of features and channels is proposed.Finally,comparison experiments are conducted on the public dataset Pascal VOC and the constructed multi-scene electronic component detection dataset,respectively.The results show that the number of parameters(13.32M)is reduced by 55% and the computational volume(28.7GMac)is reduced by 35% compared to the teacher network.The knowledge distillation method based on the combination of features and channels is able to improve the m AP of the student network by 3.91% and 1.13% on the public dataset and the multi-scene electronic component detection dataset,respectively,which is significant.(3)Multi-scene electronic component detection system implementation and analysis.Firstly,we designed the overall scheme of the system according to the functional requirements of the system,then completed the construction of the system hardware platform and the design of the software system,followed by the implementation and testing of the system functions.Finally,200 images of electronic components in different scenes of the assembly process were collected for the experimental verification of the system.The average accuracy(Precision)of the system was98.74%,the average recall(Recall)was 96.44%,and the m AP was 97.74%,which could achieve fast and accurate detection of electronic components.In summary,this thesis conducts an in-depth study on multi-scene electronic component detection methods,and provides an in-depth analysis of object detection,lightweight models,and knowledge distillation.A high-precision detection method for multi-scene electronic components is proposed.Meanwhile,in order to further realize fast and accurate detection of electronic components,a lightweight electronic component detection method based on knowledge distillation is proposed.A multi-scene electronic component detection system is designed and implemented,and relevant comparison experiments are conducted.The results show that the multi-scene electronic component detection method and the constructed detection system proposed in this study can realize fast and accurate detection of electronic components,and the research work has certain scientific and engineering significance.