Research On Scene Recognition Based On Image Semantics

Scene recognition is an important research topic in the field of computer vision.With the advent of the era of big data,many industries and fields have a greater need for scene recognition.Currently,scene recognition is mainly used in the fields of intelligent monitoring,robot vision,and disaster monitoring.Scene recognition has a history of origin,but the traditional scene recognition method has great limitations.In recent years,with the rapid development of deep learning,the research on scene recognition using convolutional neural networks has received extensive attention.In this paper,the method of deep learning is used to study the scene recognition.The main research contents include two aspects: the design of convolutional neural network and the improvement of loss function.This paper develops an efficient convolutional neural network ML-DenseNet,which is an improved convolutional neural network based on DenseNet.Compared with DenseNet,MLDenseNet mainly improves from several aspects: on the one hand,it designs an efficient dense block connection network,in the dense and fast structure,not only BN operation on the feature map after convolution,but also through the activation function.The subsequent feature map also performs BN operation to improve the convergence performance of the network.On the other hand,it predicts at multiple levels of the network,and applies two auxiliary loss functions to improve the network's utilization of shallow and middle layer feature information.Improve the feature utilization of the network;finally adopt the inverse feature fusion strategy to enhance the feature of the shallow feature,and use the deep feature map to directly enhance the shallow feature in the back propagation,making the whole fusion process more efficient.In view of the poor resolving power of the loss function Softmax Loss commonly used in current convolutional neural networks,this paper applies the loss function and introduces three efficient loss functions: LMCL,A-Softmax Loss and ArcLoss(additional angle loss function).It is verified from both theoretical and experimental aspects that these loss functions have good performance in the research of scene recognition.In order to verify the effectiveness of the scene recognition model proposed in this paper,experimental verification and analysis were carried out on two large scene recognition data sets.The Places data set is used as the training set,and tested on the test set of the Places data set and the test set of the SUN data set.In this paper,two evaluation indicators are used to evaluate the scene recognition model.The experimental results show that the proposed algorithm achieves the better recognition accuracy on the two published exhausted identification data sets,which is greatly improved compared with the benchmark data performance.