Research On Improved Feature Coding And Pooling Algorithm Based On Saliency Relationship In Image Classification

Image classification has been paid much attention and playing an important role in the field of computer vision.With more and more digital images into people’s lives,image classification technology can automatically and efficiently classify images,and has been gradually applied to various of image classification.BOW(Bag of words)model was originally used in the field of document processing,and then was applied to image classification to extract and describe the features of the image and achieved good results.Feature coding is the most important part of the BOW model.Locality-constrained linear coding(LLC)is one of the most popular feature coding methods.It creates a local coordinate for each feature descriptor by local constraint,so that similar feature descriptors can get more similar codes,and greatly improving the classification results.However,LLC doesn’t take into account the saliency relationship between features.In human visual system,the image is segmented into foreground and background parts,and the saliency value of the foreground is usually much higher than the saliency value of the background.Therefore,we believe that the saliency similar features usually contain similar information,and that the saliency relation can enhance the feature reconstruction.LLC only contains locality but ignores the saliency of the feature coding.In this paper,the research topic is based on two aspects,one is for the problem of the coding method with simple local constraints can‘t be a good representation of the image,the second is propose an improved algorithm based on the lack of contact between the feature descriptor and the visual word in the max pooling algorithm.Specific work and results are as follows.Firstly,considering the advantage of saliency information on image classification,a Locality and Saliency Similarity Constrained Coding method is proposed.This algorithm is implemented by two sub-methods: saliency similarity k NN algorithm and saliency similarity-constrained coding process.The saliency similarity k NN algorithm makes features with similar saliency value can share the same visual word by sharing the visual words of the high similarity features.The saliency similarity-constrained coding process is to improve the classification effect of the coding matrix by adding saliency similarity constraint.In this paper,the two sub-methods,each method can be run separately to improve the classification accuracy,the combination of them can further enhance the result.Secondly,saliency similarity constraints were added into the pooling process,and a saliency max pooling was proposed.In recent years,scholars have studied and improved less about the pooling process and are still the most popular of the max pooling method.However,this paper found that the improvement of the pooling process will greatly improve the image classification results.In this paper,we take into account the saliency similarity relation between the visual words and the feature descriptors,by adding saliency similarity weight to the coding matrix in the pooling process,and using the max pooling to obtain the final vector representation.It can improve the accuracy of 3%,and can also be combined with LSSC method to further improve the accuracy of classification.In this paper,through the above two research programs,the saliency similarity relation is fully integrated into the coding and pooling process,which effectively overcomes the shortcoming of the lack of contact between the features in the traditional BOW model.Experimental results show that this algorithm can not only improve the classification accuracy,but also improve the stability of encoding.