Research On Fine-grained Classification Algorithm Based On Webly Supervised

Fine-grained image classification has a wide range of applications in both industry and research,such as automatic monitoring of biodiversity,smart retail,and smart transportation,and has also had positive impacts in the fields of species conservation and production,business,etc.Fine-grained classification of species,items,commodities,etc.is critical and necessary to facilitate the development of various applications.Therefore,it is of great research value to correctly classify fine-grained images.In recent years,deep learning technology has made great achievements in the field of vision,and some researches have applied deep learning technology to fine-grained classification tasks.Due to the small differences between different fine-grained categories,deep neural networks require a large number of manually labeled data sets to train the network.Manually labeling data is time-consuming and requires professional domain knowledge.In order to get rid of the dependence of fine-grained classification tasks based on deep learning on manually labeled fine-grained datasets,this paper uses the method of webly supervision and uses easily obtained network images as the training set in fine-grained classification tasks.Since the accuracy of Internet data cannot be guaranteed,network images inevitably contain a large number of wrong labels,so label denoising of network images is very important,which can greatly improve the classification ability of the network in webly supervised fine-grained classification tasks.Fine-grained images have highly similar appearances,and the differences between images are often reflected in local areas,and images are interfered by factors such as background,pose,occlusion,and illumination.Accurately extracting discriminative local features has become a problem that affects fine-grained classification performance.key.To sum up,this paper uses deep learning as a technical means,takes the realization of fine-grained classification as the application background,adopts the webly supervised fine-grained classification method,directly uses the webly image finegrained data set as the training set,and studies the technical method of webly supervised finegrained classification.First,to solve the problem that the existing clean samples obtained based on loss still have noise and the distribution of labels generated by the network has errors,a weblysupervised fine-grained classification method based on loss and self-supervised learning is proposed.In the sample selection process,the sum of cross-entropy loss and entropy loss is used as the basis to select a small number of clean samples with high confidence;consistency loss is introduced in the label distribution learning process as an auxiliary to obtain stable and reliable internal noise samples label distribution;then introduce self-supervised learning in the network to strengthen the feature representation of the network.After improvement,the test data sets have achieved good classification accuracy.Second,fine-grained images have little difference,and there are factors such as background clutter,occlusion,and posture interference,and discriminative information is often reflected in subtle local areas,making it difficult to distinguish.This paper proposes a webly supervision fine-grained image based on attention mechanism for this problem.Granular classification method.Introduce a bottleneck attention module for the residual network,adaptively refine features in space and channel dimensions,improve the role of important features to enhance the feature network’s ability to extract local discriminative features;and introduce a skip connection strategy to convert shallow.The detailed features of the network are transmitted to the deep network,so that the deep network can obtain more discriminative detailed information.