Scene Classification Based On Convolutional Neural Network

Scene classification is the use of known semantic information to automatically annotate images on the scene library,and determine the category of images according to the content of image annotation.It is an important research direction for image understanding.In recent years,Convolutional Neural Network(CNN)has been rapidly developed in terms of speech recognition and image processing with its special structure of local weight sharing,especially in large image processing fields.Therefore,the CNN-based classification algorithm has become the mainstream algorithm for scene classification.However,on the CNN framework,the convergence speed of the network and the existence of overfitting are still one of the key issues to improve the classification accuracy of the scene.At the same time,how to improve the independence,locality,and repeatability of CNN learning features and reduce the dimension of features is also an important issue.Based on the analysis of existing literature,this paper studies the self-adaptive learning rate of CNN and the characteristics analysis through CNN learning.The main work is as follows:(1)In the framework of CNN,an algorithm that can adaptively adjust the network learning rate is proposed and applied to scene classification.CNN-based scene classification methods have achieved good results,but when the training data is small,due to the similarity and complexity of indoor scenes and other complex scenes,it is difficult to identify.It is easy to cause many times of network training,slow convergence,and overfitting.In order to eliminate this effect,the algorithm adaptively adjusts the learning rate according to the change of the error function in the network training.When the variation of the error function is small,the batch learning rate remains unchanged;when the error function increases,the learning rate changes inversely with the change of the error function.At the same time,according to the network output results,the training method of the experimental samples is changed.The training focuses on identifying inaccurate images,reducing the probability that the network will fall into an over-fitting state,and further increasing the network recognition rate.(2)On the basis of CNN learning image features,a scene classification algorithm based on Fisher feature analysis was proposed.CNN learns rich high-dimensional intermediate image descriptors through the output layer of the network,but it isInefficient to directly calculate the similarity of high-dimensional feature descriptors for image classification.In order to reduce the time of feature matching and improve the matching accuracy of similarity descriptors,the algorithm first uses CNN to supervise the training samples.Then,a low-dimensional hidden layer fine-tuning network is added between the full connection layer and the output layer to learn the low-dimensional features of the image.Then,the discriminant Fisher feature analysis is used to classify the images according to the similarity of feature descriptors between image classes,and the independence of sample features is enhanced.(3)Experiments were performed on Scene-15,Cifar-10 scene data sets and compared with current mainstream methods.The experimental results show that the proposed adaptive learning rate method improves the convergence of neural networks and effectively improves the classification accuracy,especially the complex classification accuracy of indoor scenes and other features.The image classification algorithm based on Fisher feature analysis reduces the time consumption and improves the classification accuracy.