A Visual Recognition Method For Intelligent Flexible Handling Robot Based On Deep Learning Mechanism

Part recognition is an important machine vision technology applied in industry.Accurate classification and identification of different targets in industrial environment can promote the industrial production and reduce the production costs.Currently,a part is located mostly by the fixed fixture and then identified through the image recognition technology.Because the production mode and the production task are more flexible,the traditional method of target recognition can hardly meet the requirements of flexible manufacturing diversification.According to the general recognition algorithm,the part feature is extracted by human and used to train classifier which can recognize the part.Thus,the reasonability of the designed features will directly affect the accuracy of the algorithm.Since there are various kinds of parts and complex recognition environments,the human-designed characteristics has great limitations.The feature of parts can not be expressed well to complete the recognition and classification.In recent years,deep learning is widely applied in image recognition.Compared with the traditional feature extraction algorithm,deep learning can extract feature from multiple parts by using convolutional neural network.Using deep learning to complete part recognition has better generalization ability,which can adapt to more kinds of parts,image deformation,blurring,light difference,occlusion overlap and other cases.It can better meet the requirements of flexible manufacturing.In this paper,the method of part recognition based on deep learning mechanism is carried out.The specific contents are as follows:(1)The basic theory of part recognition and deep learning is deeply studied.It is concluded that the convolution neural network has the advantages of shorter time,higher accuracy and more applicability in image recognition.At the same time,the common target detection is studied in the hope of improving the accuracy of part recognition.(2)The part recognition classification model is established.The feature map of image is extracted through the convolution neural net work to complete the task of part recognition and classification.By comparing the effects of different data and model parameters on the recognition results,it is concluded that the recognition accuracy rate of a single small part is up to 87.6%,which is higher than that of thetraditional recognition algorithm.Thus,the algorithm is proved to be effective.(3)A regional extraction network instead of a selective search algorithm is proposed.And a network for small parts recognition and detection is constructed.And thee recognition method is studied through the fusion of two networks of feature extraction and target classification.The experimental results show that the accuracy rate of location detection is 90.73%,the recognition time is 78 ms.And the recognition rate for complex environment is also 87.5% with 86 ms recognition time.Compared with the common target detection algorithm,this method is proved to be real-time and can meet the needs of industrial production detection.