Visual Classification And Recognition Based On Deep Neural Network

Vision is the cornerstone of robot intelligence,and image recognition is based on vision.SIFT+SVM and HOG+SVM algorithms are applied in the field of robot image recognition.The recognition rate is difficult to break through all the year round in that the limitation of the algorithm.With the development of artificial intelligence,deep learning is becoming more and more serious,especially the convolution neural network has made a great breakthrough in the field of computer image,so it is more practical to integrate the deep neural network and the actual situation of the robot sorting task,and to improve the intelligence of robot sorting by the depth learning.This paper analyzes the defects of traditional image recognition algorithm,reports the research status of deep learning at home and abroad,and determines how to use convolution neural network to improve the efficiency of robot identification and sorting.This paper introduces the basic knowledge of the deep neural network,analyzes the network structure of the convolution neural network in detail,and deduces the forward and back propagation algorithm of the convolution neural network,which provides the theoretical basis for the establishment of the image recognition model.In robot sorting and recognition,there are many tasks to identify the quality of the same object.In order to satisfy this fact,the paper created the Apple DataSet.The dataset contains 3926 images,which are filmed on the spot based on the actual recognition task.It is more practical than common mnist and CIFAR-10 datasets.This paper aims at the problem that the robot has low efficiency on the same object quality,an Apple quality recognition model based on depth neural network is designed,which is composed of input layer,6-layer convolution layer,2-layer full connection layer and output layer.In order to obtain better and more specific image features,in the design of the third convolution layer,The principle of Hebbian is introduced,which is able to connect some highly correlated features in the same spatial location but different channels.The paper also expounds the design concept of other layers,and calculates the parameters that need to be trained in each layer.The code of image preprocessing is designed by using Python language;According to the calculation parameters of network recognition model,the Apple quality identification model is built on the framework of TensorFlow deep learning;In order to avoid gradient dispersion,a relu activation function with unilateral inhibition is used;In order to speed up the convergence of the cost function,The stochastic gradient descent method is used to train the network.Through training,some important parameters on the performance of the law are obtained.Finally,the training model is tested,and the experimental results show that the model can be used to identify the quality of the same object very effectively.Compared with the traditional recognition algorithm,the recognition model is morethan 95% correct for Apple quality recognition,and the recognition time is shortened greatly.