Research On Visual Features Extraction Of Industrial Robot Based On Deep Learning

Object recognition and classification is always play an important role in automatic sorting,moving and testing system acted by industrial robot,as well as the foundation of object tracking,behavior analysis and scenes understanding.Using traditional feature extraction method to get image feature and training classification to get classification model is the widely used solution of object recognition and classification at present.However,feature extraction method designed by prior knowledge may affect classification accuracy directly.Convolutional neural network can extract mid-level features and also overcomes the influence of image deformation,Occlusion and illumination variation,which has wide applications.However,there are too many parameters to learn,which may lead to time consuming and overfitting.This paper study on the problems of classification by industrial robot,specific tasks as follows:(1)We have collected data and literatures of this subject and made overview of visual problem in automatic sorting,moving and testing system acted by industrial robot.Then,made overview of basic visual feature extraction and feature representation in classification problem acted by industrial robot.(2)We introduced the structure of neural network as well as its advantages in classification tasks,especially with the emphasis on hierarchical structure in convolutional neural network and the concrete process in classification.(3)In accordance with problems such as too many parameters,too long training time and too skillful adjusting parameters in traditional convolutional neural network,we proposed a simplified deep learning model called Locally Linear Embedding Network.LLENet extracts convolution kernels by locally linear embedding algorithm instead of back propagation algorithm,which generating complicated calculation used in traditional convolutional neural network.Similar to traditional convolutional neural network,LLENet cascade convolution layer,down sampling layer,non-linear layer to extract features and training Support Vector Machines or K-Nearest Neighbor classifications.In this paper,we do experiment on MATLAB R2012a.Firstly,experiments carried on WorkPiece dataset which collected in industrial robot field of view shows that classification accuracy of LLENet is higher than PCANet and KPCANet,where LLENet with one convolutional layer can reached 97%accuracy that outclass PCANet and KPCANet.Secondly,training LeNet-5 convolutional neural network with WorkPiece dataset using Caffe framework,which achieve over 95%accuracy.Last but not least,fine-tuning on LeNet-5 convolutional neural network with WorkPiece-1 dataset,which is the subset of WorkPiece to simulated lacking of training data that achieved over 93%accuracy.The results of experiment show that LLENet has better classification performance,lower complexity than traditional convolutional neural network,which helps to improve accuracy of industrial robot classification.Fine-tuning on pre-trained LeNet-5 can also generated good classification model when lacking of training dataset.