Design And Implementation Of Sorting System Based On Deep Learning Of Machine Vision

Automated workpiece sorting is an important basis for smart factories.It improves the industrial productivity and solves the problem of manual inoperability in certain environments.The application of machine learning techmology in sorting system is the inevitable trend of development in industrial assembly line.Convolutional neural network,as one of the machine learning technologies,Mth its excellent performance and great potential,it has attracted many industry scholars and experts to participate in the research.At present,the recognition and localization of target objects based on convolutional neural networks has become a research hot spot in the field of object detection.In this thesis,we take mechanical parts of industry as the research object,and take the industrial conveyor belt as the research scene,focusing on an improved YOLOv2 convolutional neural network.The specific contents are as follows:1.the improved model modifies the activation function of YOL0v2,The Se ReLU activation function is designed.The new function has both the sparsity and convergence of the ReLU fimction and the smoothness of the Sigmoid ftmction.Even the convolution result is zero,it can still to perform nornal training.2.Replace the third and fourth large layers of the original network with the Inception model.The modified architecture reduces the parameters of trailing while improving the performance of network,and improving the utilization of computing resources.3.Multi-scale and pre-training methods of network are used in the new network.Multi-scale training makes the network robust to pictures of different input sizes,an pre-training improves the accuracy of object recognition while reducing network training.time.4.In the new network,the k-means dimension clustering algorithm is used to cluster the target frames manually marked in the dataset,and the statistical rules of the target frames are found,and the candidate frames that meet the specific detection tasks are obtained.By adopting this method,on the one hand,the accuracy of target positioning is improved,on the other hand,the number of candidate frames is reduced,and the parameters of training are reduced.Then,the improved network is trained on the data sets we collected,and the experimental data is analyzed and summarized in the three aspects of average recognition accuracy,detection speed and convergence speed of network.The experimental results show that the improved convolutional neural network is superior to the original YOLOv2 network in terms of recognition performance and time efficiency.Finally,we use the trained network for our own mechanical parts sorting system.In the complex scene with interfering objects,the average accuracy of recognition is 85.8%,and the detected speed reaches 25 frames in per second,which laids the foundation for the real-time sorting of workpieces on industrial assembly lines.