Research On Image Classification Algorithm Based On Visual Confusion Characteristics

In recent years,image classification algorithms have developed rapidly and achieved remarkable results.Especially,the image classification algorithm based on deep learning has achieved the performance of human shoulders.However,there are still two major challenges in the development of this field.First of all,although the image classification algorithm based on deep learning performs well,the current mainstream algorithms do not make full use of a priori information such as visual confusing.Secondly,the volume and computational complexity of the deep learning model are very large and difficult to meet.Real-time application requirements for end devices that have limited power and capacity.In this paper,the two challenges faced by image classification algorithms are introduced,and the visual confusion characteristics that are common in images are introduced.Firstly,in order to characterize the visual obfuscation characteristics of images,we have established a tree structure of “visual confounding trees”for image datasets.Through the “visual confusion tree”,each category in the image dataset can be classified into different granularities according to the hierarchical structure.A collection of categories that reflect different levels of confusion.Secondly,based on the visual confusion tree,a tag tree classifier and a tag tree classifier with backtracking are established.The visual obfuscation tree is combined with the traditional machine learning method to increase the accuracy of image classification.At the same time,the visual confusion tree is also used.The structure is embedded into the depth model to enhance the performance of the image classification.When the visual obfuscation tag tree is established,the tree classifier can be used to replace the fully connected layer of the depth model with a very large amount of computation,so that the real-time performance of the depth model image classification algorithm can be enhanced.We first verified in our experiments that our tag tree classifier has been significantly improved over the current best performing tag tree classifiers,with Top-1 precision in the CIFAR-100 and ILSVRC12 data sets respectively.Increased by 4.3% and 2.4%.In addition,our approach has a speedup of 124 and a cost savings of 115 compared to the AlexNet and VGG16 models with fully connected layers,respectively,without sacrificing accuracy.Then,the effectiveness of our backtrackable tag tree classifier algorithm is verified by experiments.Experiments are carried out on the CIFAR-100 data set.The experimental results show that the features extracted by different depth learning models,our backtrackable tag tree classification The accuracy of the tag tree classifier is higher than that of the tag tree classifier without backtracking.Finally,we verify the advantages of our proposed method by comparing the proposed visual tree convolutional neural network with the reference convolutional neural network.In the experiment,we constructed 3 different visual tree convolutional neural network models.We found that the 3 visual tree convolutional neural network models have accuracy in comparison with the corresponding baseline deep convolutional neural networks.An increase of 1.36%,0.89%,and 0.64%.