Research On Detection Technology Of Aliasing Electronic Components Based On Embedded Deep Learning

With the development of computer technology,object detection technology has been widely used in industrial robots and industrial manufacture.However,the detection accuracy and speed of small objects under complex background in the manufacture environment still cannot meet the manufacture requirements.In addition,at present,most detection algorithms need to run in industrial computers with strong performance.For embedded devices with weak performance but more advantages in volume and power consumption,detection algorithms are still unable to realize rapid detection on such devices.In order to solve the above problem,this topic on the National Natural Science Foundation of China "Research on Vision Autonomous Recognition,High Precision Positioning and Compliance Control Method of Intelligent Assembly Robot",in view of the aliasing electronic components detection algorithm,based on comparing the existing detection algorithm and selection for small object detection is the most dominant YOLO-V3 detection algorithm,and two step improvement is carried on.The improved detection algorithm can realize rapid detection in embedded devices on the premise of ensuring detection accuracy,which lays a foundation for the function realization of subsequent intelligent assembly robots.This thesis mainly includes the following research contents:1.Introduces the hardware system needed to build the dataset of aliasing electronic components,including the selection of camera and light source,and the performance indicators of training and deployment platform,etc.The advantages and disadvantages of the existing object detection algorithms are analyzed.For the aliasing electronic components in this subject,the YOLO-V3 algorithm which is most suitable for the detected object is selected to construct the detection algorithm.Finally,the evaluation index of detection algorithm performance is introduced to provide the basis for the comparison of detection algorithm performance.2.Aiming at the detection accuracy and speed of aliased electronic components,an improved algorithm of YOLO-V3 based on Mobilenet network(called YOLOV3-Mobilenet)is proposed.By replacing the standard convolution in the YOLO-V3 backbone network with the deep separable convolution and adjusting the network parameters,the speed of the detection algorithm is improved.At the same time,image enhancement algorithm and Mixup method are used to optimize the training process of image data and detection model to further improve the accuracy of detection model.Experiments show that the YOLOV3-Mobilenet detection algorithm,compared with the YOLO-V3 algorithm,doubles the detection speed on the condition that the accuracy is basically the same,achieving a balance between detection speed and accuracy.3.To solve the problem that deep learning detection algorithm is applied to embedded devices with low operating efficiency,a sensitivity based YOLOV3-Mobilenet convolutional kernel clipping algorithm is proposed.By analyzing the sensitivity of each convolutional layer in the YOLOV3-Mobilenet detection model,the optimal clipping rate of the detection model was obtained,and the convolutional kernel clipping of the convolutional layer in the YOLO-V3 detection head in the detection model was performed with this data.Finally,the size of the detection model was reduced by 87% when the detection accuracy was slightly reduced.The detection speed is increased to 1.78 times of the original model,which meets the deployment requirements of embedded devices.4.Build the software system in the detection system of aliased electronic components,including the design of interactive interface and functional modules,and complete relevant tests to verify the feasibility and effectiveness of the detection algorithm of aliased electronic components based on embedded deep learning in this paper.The experimental results show that the proposed algorithm can meet the requirements of various indicators of the detection of aliased electronic components,and is feasible and efficient.To sum up,this thesis conducts research of the detection algorithm of aliasing electronic components based on embedded deep learning.And the balance between detection speed and accuracy and the performance loss of embedded device deployment are discussed in detail.In this thesis,a stable and efficient embedded detection algorithm is proposed,and a hybrid electronic component detection system is established.The intelligent and lightweight electronic detection system is realized,which lays a solid foundation for the function realization of intelligent assembly robot.