Research On Scene Image Classification Algrithm Based On Deep Learning

In recent years,with the development of computer technology and artificial intelligence,deep learning has made tremendous achievements in the field of computer vision.Scene recognition,as one of the important research directions in computer vision,has larger application value,and has a wide range of applications in intelligent robots,automatic driving,image retrieval and so on.However,the composition of scene images is complex,usually including multiple obj ects,and the same obj ect will exist in different scene images,which makes the scene recognition challenging.In addition,because of the different illumination,angle and so on,the same kind of scene usually presents different shapes,which also brings difficulty to scene recognition.Traditional scene recognition methods use shallow features to classify scenes.Shallow features can not express the global information of scene images and the semantic association between obj ects in images.Features extracted based on deep learning algorithm can well represent the local and global information of scene images,which greatly improves the accuracy of scene recognition.However,the algorithm based on deep learning can not express the global information of scene images.Deep convolution neural network consumes a lot of computing resources.At the same time,the differences between classes and the diversity within classes of scene images are still difficult problems in recognition tasks.It is difficult for common object classification networks to completely solve their impact.Based on this,this thesis the convolutional neural network is researched,the network structure and classifier of the convolutional neural network are improved,and a scene recognition algorithm combining multi-resolution and multi-scale is proposed.The main contents of this paper are as follows.(1)The structure of convolution neural network is studied in this paper,by combining the advantages of Inception structure and Short-Cut structure,and introducing deep separable convolution structure to optimize the operation speed,the network complexity is reduced and the efficiency of the model is improved.The proposed improved convolution neural network is tested on the scene image data set.The results show that compared with the current popular convolution neural network,the proposed improved model can effectively reduce the running time of the model while guaranteeing the classification accuracy.(2)In order to solve the problem of high similarity and complexity of scene images,the convolutional neural network classifier is studied.L-Softmax algorithm for face recognition is introduced into the scene recognition,which enhances the distance between classes while training scene images,and avoids over-fitting.Experiments show that the improved algorithm achieves 95.06%recognition accuracy on scene 15 data set,far exceeding the traditional one.The feature extraction algoritlhm is superior to the general scene deep learning algorithm.(3)In order to adapt to the multi-scale characteristics of scene images,a multi-scale and multi-resolution network model is proposed,which uses multi-resolution to learn the local and global features of scene images,and uses multi-scale to optimize the impact of different scale objects in scene images on recognition tasks.The combination of multi-scale and multi-resolution models can fully express the information of scene images,and the recognition accuracy is much higher than that based on traditional shallow feature extraction methods.At the same time,the proposed network structure can achieve end-to-end training and prediction,with higher flexibility and lower complexity.