Deep Learning Based Sorting Robot Target Recognition And Location

With the development of science and technology,various industrial robots have been widely used in industrial production,daily life and other fields.As one of the important branches of industrial robots,sorting robots can not only improve the production efficiency,but also reduce the production cost when the labor cost is high.The machine vision system is introduced to the sorting robot,and the recognition and positioning of the sorting robot vision system in the complex environment is the target detection task,which is the technical basis of the automation of the sorting robot.The traditional target detection algorithm needs to extract the target features manually,and then input the extracted features into the classifier for training classification.However,in the multi-target complex environment,it is very difficult to extract the feature operator manually.In recent years,with the rise of deep learning,especially the convolutional neural network framework has shown great advantages in image processing tasks--the target detection algorithm using convolutional neural network as the basic architecture has better generalization.This algorithm can well adapt to the influence of multi-target detection,weak light and other complex environments,and better meet the requirements of intelligent sorting robot vision system.In this paper,'the target detection algorithm based on convolutional neural network is studied,and the specific work contents are as follows:(1)the process and omission of traditional target detection algorithm are briefly described,and the structure,training mode and related algorithm of convolutional neural network,especially the theory of deep learning,are studied.(2)three models of RCNN series algorithm are introduced.Aiming at the problem of different types of fruit recognition,a target detection algorithm model based on regional convolutional neural network is established.The validity of the algorithm for multi-type fruit detection is verified by experiments.At the same time,the RCNN algorithm and FasterRCNN algorithm are compared.It is verified that the FasterRCNN algorithm not only has a much higher detection speed than the RCNN algorithm,but also has a higher target detection accuracy.(3)the reasons for the high detection error rate when oranges and apples appear at the same time are analyzed.The following two improvements are made:first,the structure of feature extraction network is changed and a new feature extraction network model is established.The training parameter size of the network model is about 0.21MB,which greatly reduces the storage space of the network model in the computer.The cifar-10 training set is used to verify that the performance of this network model is better than that of other lightweight network models.In the target detection task,the accuracy of using this network model as feature extraction network is 3 percentage points higher than that of using ZFNet as feature extraction network.Secondly,aiming at the problem of too few training sets,the image preprocessing method for some training set images is used to realize the expansion of training sets,so that the network model has better generalization.The experiment proved that the accuracy increased from 75.3%to 83.52%.(4)target detection algorithm is combined with traditional image processing to locate the capture point of the target.For some fruits with special shapes,the location of the mark box in the target detection algorithm is not very accurate,and the use of the center of the mark box as the grabbing point of the target may have a large error with the actual situation.In view of this situation,this paper first extracts the detected target,then uses k-means algorithm and otsu method for image segmentation and binarization,and finally finds out the location of the center point.The feasibility of this method was verified by experiments on fruits in different positions.