Research And Application Of Image Classification Based On Multi-Cue Feature Fusion

Image classification is one of the research hotspots in the field of computer vision,and it is also the basic stage of image processing research.The image classification algorithms proposed by many researchers mainly depend on the extraction techniques of image features and the construction methods of classifiers.The methods commonly used to extract image features are primarily based on low-level visual features of the image,such as the color,texture,and shape of the image.Support Vector Machine has supervised learning ability and powerful generalization ability,which is one of the first choices for most scholars to study image classification.Although there are many image classification algorithms based on multi-feature fusion,there are still some problems be solved,such as the presence of noise and redundant information in the image,resulting in the extracted features not well characterized by image information.When BT-SVM multi-class classifier solves the classification problem,it is easy to construct the "error accumulation" phenomenon,resulting in low classification accuracy.Aiming at the above problems,this paper proposes an image classification algorithm based on multi-cue feature fusion based on a large number of literatures and related data,and applies the algorithm to flower image classification.The main research contents of this paper are as follows:(1)For the multi-feature fusion method,the image information can not be well characterized,and a multi-cue feature fusion method is proposed.First,by compressing the original image to generate a compressed image,this compressed image can reduce redundant information in the image.The final saliency map is obtained by the fusion of the significant graphs generated by the global and the rarity.In terms of global saliency,the global saliency of the three color channels L,a,b is fused by the weighted linear method,and the weight is based on each channel.The ratio of the average significant value to the sum of the average significant values of the three color channels is generated.In terms of rarity,the saliency of pixels that occur less frequently in the image is set to zero.Intensify the fusion of the two metric saliency maps,adding the average saliency value as the adjustment factor,and widening the gap between the image foreground and the background significant value.Then,by extracting the HOG features in the three images,and finally performing the stitching fusion,the blended feature vectors can better represent the image information.(2)For the OVO-SVM classifier and OVR-SVM classifier,it is necessary to construct more number of classifiers,and BT-SVM multi-class classifier will cause "error accumulation" phenomenon.This paper proposes a DBT-SVM multi-class Classifier.The DBT-SVM multi-class classifier adopts the generation strategy of approximate complete binary tree and the class distance method in clustering,which not only can achieve the optimal complete state of the structure of the binary tree,but also can preferentially separate the most easily separated categories.(3)The algorithm was validated on the Caltech101 and 17 Category Flower datasets.Through comparison of different experimental results,the proposed algorithm can effectively improve the classification accuracy,and it also has certain advantages in the application of flower image classification.