The Study Of Scene Classification Algorithm Based On The Combination Of Global And Local Features

As the saying goes: "The eyes are the windows of the soul".Visual is the main way for human being to perceive external information.Therefore,in the computer field,machine vision and image understanding have been the research hotspots.Machine vision has a lot of applications in real life.For example,a vision-conscious robot firstly determines the scene in which it locates,and then can recognize the surrounding objects or make a response to the environment.As the information society continues to evolve,a large number of pictures are stored on the smart devices on the Web and in people's hands.These photos can be categorized,labeled and retrieved according to different scenarios.Based on these potential and real needs,in 2006 year,technical seminar on Scene Recognition and Understanding at the Massachusetts Institute of Technology identified scene recognition and understanding as a new and promising research arm in the field of machine vision.This paper studies and improves the scene classification and recognition algorithm based on gist features.The word of Gist in the dictionary means "the subject," as its name suggests that gist is a feature that reflects the contextual context of a scene.The Gist feature model is a feature obtained by sparsely meshing a feature image of an image or an image.As a result,the gist feature only reflects the global features of the image,and its granularity is too coarse.In scene recognition,the local edges and textures in the image are also helpful for the successful scene recognition.Therefore,in this paper,we can combine the global histogram with the gradient direction histogram that can describe the local edge and texture information of the image,and propose the gist feature based on the multi-scale gradient direction histogram.Histogram of Oriented Gradient(HOG)is an excellent local edge and texture description algorithm that is insensitive to small offsets and changes in illumination in the image.In order to describe the contour and texture features of different scales in the scene picture,this paper attempts to use the gradient direction histogram and the multi-layer gradient direction histogram(separately extract the gradient direction histograms on different scales of the scene picture)to extract the local edge and outline information of the scene picture as a complement to gist features.The main work of this paper is as follows:In this paper,we use the famous gist feature extraction model proposed by Itti as an improved model.Firstly,the gist feature extraction method based on multi-layer gradient direction histogram is proposed.The method uses the multi-layer gradient direction histogram to extract the local contour and texture features of the image and makes a simple concatenation with the gist feature proposed by Itti.The aim is to supplement the local contour and texture information of gist features with coarse grain size.After that,we use the multi-layer gradient histogram to improve the directional feature extraction channel of gist feature extraction model proposed by Itti.In this paper,the improved multi-layer gradient histogram replaces the four-angle and four-direction Gabor filters in the original feature extraction model as the direction feature extraction algorithm in the gist model.The multi-level gradient histogram has better local contour and texture description ability than the four-angle and four-direction gabor filters,which can make up for the coarse particle size of gist feature extraction model proposed by Itti.Moreover,the second improved method has a shorter eigenvector length than the eigenvector obtained by the first improved method,so the computational cost will be lower.In this paper,the improved model is called gist feature extraction model based on multi-layer gradient direction histogram.Finally,this paper uses MATLAB to achieve the improvements proposed above for the gist feature extraction model,and carries out several simulation experiments on the OT scene recognition experimental database,and compares the performance of the model between the two improved methods proposed in this paper and the gist feature extraction proposed by Itti.Experimental results show the effectiveness of the proposed method.