Research And Application Of Image Classification Based On Local Features

With the rapid advancement of science and technology,image classification technology has become a hot topic in the field of computer vision.Among them,local feature-based image classification has been widely used in various fields.However,local feature algorithms lack sufficient generality,and corresponding adjustments and improvements are needed when applied in different fields.In this paper,through the research and analysis of local feature-based image classification technology,we propose the RM-SURF algorithm based on image retrieval and matching,and apply it to the classification of bullet shells,achieving the traceability of bullets.Due to the difficulty in large-scale collection of bullet shell image data and the scarcity of training data,this paper first studied data augmentation techniques for bullet shell images.To address the issue of incomplete,deformed,or fragmented bullet shells in practical applications,the following methods were used for data augmentation: on the basis of the original image,the target image was randomly cropped to randomly generate missing parts in the bullet shell position,achieving data augmentation of incomplete and fragmented bullet shell images.By analyzing the material properties of brass bullet shells and using ANSYS simulation tools to conduct simulation experiments,the size of bullet shell deformation under different pressures was obtained.Experimental verification of bullet shells was performed using pressure experiments.Based on this,combined with sine wave curves and image liquefaction technology,data augmentation of bullet shell deformation images was achieved.Conventional techniques for data augmentation of bullet shell images were also used,such as adding noise and rotation.This enriched the training data set and laid the foundation for subsequent research work.Furthermore,this article combines the theories of image retrieval,image matching,and the SURF algorithm,and proposes the RM-SURF algorithm,a local feature-based image classification method.This method addresses the issue of erroneous matching in the SURF algorithm by incorporating the characteristics of seashell images,introducing an adaptive threshold to constrain the ratio of nearest neighbors in both directions,and combining similarity comparison techniques to achieve seashell image classification.To address the issue of unstable classification accuracy of tiny seashell fragment images in the RM-SURF algorithm,the cosine constraint method is introduced to further improve the similarity comparison technique and achieve stable accuracy.Experimental results show that in the matching stage,the matching accuracy of the RM-SURF algorithm is improved by 2.8% compared to the traditional SURF algorithm;in the classification stage,the classification accuracy of the RM-SURF algorithm is improved by 17.1% compared to traditional classification algorithms.In practical engineering applications,the classification results of this article are input data for subsequent work modules,and the proportion of positive samples retained in the classification results is an important factor that affects subsequent work.Therefore,this article proposes the K-retention rate evaluation index,which is the proportion of positive samples contained in the top K classes of images with the highest similarity.Through experiments,it is found that the retention rate of the RM-SURF algorithm is 97.9%(K=3),which meets the needs of practical engineering applications.