Research On Image Classification Algorithm Based On The Bag-of-words Model

With the rapid development of mobile Internet,the number of digital image has been increasing rapidly.In the face of massive amounts of image data,how to extract the effective information from image data and manage them uniformly has become one of the research hotspots in the field of computer vision.Therefore,the technology of image classification can help users quickly complete images analysis and management,greatly improving work efficiency.In the research of image classification,bag of words model is widely applied.Through the in-depth study of the bag of words model in image classification,the thesis proposes an improved method to solve the problems such as the large feature encoding error and the inaccurate understanding of image content in the model,and further improve the accuracy of image classification.The thesis proposes improvement methods on feature coding and feature expression.The main research work is as follows:(1)In order to enable the image content to be expressed accurately after feature encoding,the loss of image information is reduced,thereby improving the accuracy of image classification.Based on fast low rank image classification algorithm,a coding method to enhance the local constraints is proposed.In the coding process,the features in the image are first clustered to obtain a set of local similar features and their corresponding cluster centers.Then,the nearest neighbor strategy is used in the visual words to find K neighboring visual words corresponding to the cluster centers and compose the corresponding visual dictionary.Finally,the fast low-rank coding algorithm is used to encode the feature,which makes the local similar feature root have similar coding,reduces the reconstruction error of the feature and accurately expresses the image content.(2)When the word bag model is used to classify image,the background information of the images is complex,which causes certain interference to the feature extraction,which makes it difficult to extract discriminative features in the image,resulting in failure to completely express the image content.To solve this problem,multi feature extraction is used to describe the image,avoiding the limitation of single feature to image description.At the same time,in order to obtain better feature expression,the feature can be effectively fused,combined with multi kernel learning algorithm to get the weight parameters of the kernel function corresponding to different features.Then different kernel functions are combined to get the final combinatorial kernel function.The multiple features can be better combined to enhance image expression,thereby improving the accuracy of image classification.In this thesis,several groups of experiments on scene and object image database are carried out.The results show that the improved algorithm helps to enhance the discriminability and integrity of the expression information of the image features,and further improve the accuracy of the image classification.