Research On Image Feature Generation Method With Scale Space

Since the 21 st century,with the rapid development of the mobile Internet,as an important carrier of information dissemination,image has become faster and faster.Therefore,how to efficiently process and identify images effectively becomes an urgent problem to be solved in the field of computer vision.Just as the human eye recognizes the actual image,understanding the image makes it possible to recognize the image.Therefore,in the field of computer vision,how to understand and describe images reasonably and fully has become the main research topic nowadays for many scholars in this field.Scale space,as an important representation of image space,has long been an important method and means for scholars to represent images because of its high consistency with human retina perception of real things.The traditional method of image feature generation based on scale space is more to describe the image space information.In addition,the image also includes the global and local shallow visual features and deep semantics such as content and emotion.In addition,the past scale space division depends more on expert experience,which is not conducive to show the reliability and scalability of the algorithm.In view of the above problems,this paper introduces the shallow visual information of the original image into the scale space and uses the adaptive strategy to divide the scale space.Finally,the original image is expressed more fully and effectively.The main contents of this paper are as follows:(1)Aiming at the problem of traditional visual pyramid matching model-based feature extraction algorithm ignoring the shallow visual information of image,an image feature generation algorithm based on visual information of superficial image is proposed.Firstly,the original image is divided into the gradually faded sub-regions.Secondly,the corresponding texture features are generated based on the gray-level co-occurrence matrix for the sub-regions at different levels,and the weight of each sub-region is determined through the adaptive query and fusion strategy.Finally,the resulting texture features are integrated into the SIFT features to get the final feature representation of the target image.The obtained features are used for image classification tasks,and experiments are performed on the datasets 15-Scenes,Caltech101,and Caltech256.The experimental results show that compared with other related feature extraction algorithms,the image can be more effectively described with the connection of spatialinformation and shallow texture information.(2)In view of the fact that the existing scale space division method relies more on the expert experience,an adaptive scale space division algorithm based on particle swarm optimization is proposed.Firstly,Particle Swarm Optimization(PSO)algorithm is used to optimize the maximum and minimum spatial scale partition ratio.Secondly,the cumulative stability evaluation method is incorporated into the PSO framework.Finally,the image features obtained are used for image matching tasks.The comparative experimental results on Oxford and Fischer image matching datasets show that the image can be described and expressed more effectively by finding the optimal spatial scale partitioning scheme adaptively.(3)Most feature generation methods based on scale space ignore the focus of attention.In addition,the existing methods often give the same importance measure to different concerns,and propose a method of image feature generation based on multi-pyramid space structure.Firstly,the original image is divided into the gradually faded sub-regions,and the number of feature points detected by the detection operator is used as the importance measure of each sub-region.Secondly,for each sub-region,the corresponding scale space is generated with Gaussian smoothing kernel function,and the cumulative stability evaluation method is used to generate the corresponding features.Finally,the clustering algorithm is designed to parallelize the computing framework to reduce the time consumption of the algorithm.The generated features are used for image matching tasks.The comparative experiments on Oxford and Fischer image matching datasets show that the parallel image features generation method based on multi-pyramid spatial structure can generate image features more quickly and efficiently.