Research On Image Recognition Robustness Based On Depth Learning

In recent years,due to the excellent performance of deep learning in image recognition,it has been widely studied and applied.Compared with shallow learning,deep learning changes the lowlevel features layer by layer and extracts high-level features,which makes deep learning have stronger classification ability.In the existing research,researchers pay more attention to how to improve the recognition ability of the algorithm,but less about the robustness of the algorithm.However the robustness of the algorithm is one of the most important criteria of an algorithm.In order to enhance the robustness and improve the application range of deep learning algorithm,the paper mainly does the following work:(1)In order to reduce the negative impact on the training of neural network when the error of tagging the sample tag,this paper presents a pre-training tag pre-judgment algorithm.In practical applications,the data samples for deep learning are labeled manually,but the labels may be marked incorrectly.The algorithm proposed in this paper judges whether to modify the tag of sample before training and then train the network to reduce the wrong label samples.Experiments show that the algorithm can effectively reduce the harm of false labels to neural networks and improve the robustness of deep learning image recognition.(2)A sample training method is proposed,which is different from the Minibatch stochastic gradient descent method.After the neural network has been trained for some time and has the ability to classify the images.the current classification ability of the neural network combine the information entropy are used to divided image samples into simple samples and confused samples,and put them into the simple sample set and the confused sample set respectively.In the subsequent training,the confused samples and the simple samples are specially trained for a certain number of iterations.In this way,the neural network will further study features of different types that distinguish with each other,thereby improving the ability of the algorithm to classify images.Experiments show that with this algorithm,neural network can effectively reduce over-fitting and improve the robustness and generalization ability of the algorithm.(3)Finally,this paper applies two algorithms in image recognition system and implements a commodity recognition system.In this system,the user can upload the commodity images to the system through the client for training.At the same time,the user can also upload the commodity image to the system through the client for identification.Then the system returns the name of the product to the user after recognizing the image uploaded by the user.Experiments show that the system has some practicality.