Research On Target Sorting Algorithm Based On Machine Vision

With the continuous progress of social science and technology,the target sorting based on machine vision has made great progress in industry,agriculture,medical treatment and other fields.With its fast,efficient and uninterrupted working characteristics,the target sorting method based on machine vision has gradually replaced manual sorting and traditional machine sorting methods.Due to different industry characteristics,there are significant differences in the target objects being sorted.Traditional machine vision sorting algorithm with a hand-designed feature extractor,When the feature extraction of the target object is performed,the deep feature of the target object cannot be extracted.When the detected target has complex features or the sorting environment is complex,the traditional target sorting algorithm is difficult to achieve target sorting.The object sorting system based on machine vision mainly includes three parts: image acquisition,object detection and sorting.Target detection is the core of the whole system.In this thesis,Faster-RCNN target sorting algorithm based on deep learning is selected as the target detection part of the whole system.It can be used to deep learn the features of object and extract the features of object in depth automatically,so as to make up for the shortcomings of traditional target sorting algorithm.The shortcomings of depthing learning target sorting algorithm are improved in terms of detection speed and detection accuracy.The algorithm is validated on the filtered public data sets,and the actual sorting experiment is carried out on the self-built two kinds of data sets combined with Baxter robot.The results show that the target detection algorithm based on deep learning can successfully detect different kinds of target objects in complex scenes and achieve sorting,with high accuracy.Based on the current main Faster-RCNN target sorting algorithm,the feature extraction part is improved in two aspects to improve the detection performance of the algorithm.The main research contents are as follows:(1)By modifying the form of feature extraction network convolution core in Faster-RCNN algorithm,the network detection speed is improved: Faster-RCNN detection algorithm based on convolution neural network,feature extraction part as the main module of the whole detection algorithm,its forward propagation speed and feature extraction ability,affect the performance of the whole network.In order to reduce the network parameters and improve the speed of feature extraction,the advantages of ResNet network and Mobilnet network are combined.On the basis of retaining the advantages of the residual module,the standard convolution core in the residual module is replaced by the deeper separable convolution core with smaller parameters in the MobileNet network,so that the network can maintain high accuracy and improve the network detection speed.(2)By modifying the number of convolution kernels in Faster-RCNN feature extraction module,the network width is increased and the accuracy of network detection is improved: By comparing the structure of Inception network,it is found that the Inception module uses several convolution cores of different sizes to extract features from feature maps,which increases the adaptability of the network to the scale of feature maps,obtains more spatial information and improves the ability of feature extraction.In this thesis,a single separable convolution core in the residual module composed of the improved single-core deep separable convolution is replaced by three deep separable convolution kernels: 3?3,5?5,7 ?7 and the feature map are connected as output.The improved classification network is replaced by the feature extraction part of Faster-RCNN to improve the accuracy of detection algorithm.(3)The improved detection algorithm is validated by establishing different kinds of data sets.The data sets of jujube and walnut related to agriculture were established.the data sets of screw,nut,right angle iron related to industry were established and automobile and aircraft models three types of data sets.The sorting experiments on two kinds of data sets are carried out using the improved detection algorithm,which verifies that the target sorting algorithm based on the improved depth learning can make the machine vision sorting system have stronger sorting ability in complex scenes.At the same time,it is proved that the improved algorithm has better adaptability in different kinds of target detection.