Fine-grained Image Classification Based On Feature Selection And Multi-scale Feature Fusion

In order to further improve production efficiency and reduce labor cost,a large number of enterprises gradually pay more attention to the development of related products in the field of artificial intelligence.The computer vision as an important part of artificial intelligence has attracted more and more researchers' attention,especially the direction of fine-grained image classification has wide application prospect.If low-cost fine-grained image classification technology can be developed and applied to the monitoring of the ecological environment,the development of ecosystem research can be further promoted.In addition,it can also improve the efficiency of self-checkout system in the supermarket and realize the effective tracking of suspected vehicles in traffic accidents.However,fine-grained image classification become a challenging research direction because of intra-class variance and inter-class similarity.Only those with professional knowledge can effectively identify different subclasses,so the cost of marking is too high.In order to overcome these difficulties in the fine-grained image classification tasks,researchers have developed a large number of fine-grained image classification algorithms.These algorithms can be roughly divided into three categories,it includes fine-grained image classification based on part localization and fine-grained image classification based on fine-grained feature learning and fine-grained image classification based on visual attention.These algorithms are similar in that they learn the fine-grained features in the discriminative regions.Only by grasping these finer features that are easy to distinguish from other subclasses can we effectively classify different fine-grained images.Therefore,our research goal is to achieve effective classification of different subclasses of fine-grained images without extra bounding box or part annotations.To achieve this goal,we need to localize the discriminative regions and learn finer features in the image.In this thesis,we explore fine-grained image classification methods depend on image-level labels.The main works are as follows:(1)Aiming at intra-class variance and inter-class similarity in the fine-grained datasets,we propose a fine-grained image classification method based on feature selection.This method improves the quality of the feature map by integrating the effective information of different receptive fields.In order to select discriminative features,part regions of different scales are scored and finer features in the discriminative regions are extracted for classification.The proposed loss function is used to help our model to localize discriminative features,which not only solves the problem of intra-class variation,but also prevents overfitting in the particular samples,so that it solves the problem of inter-class similarity.We conduct experiments on three commonly used fine-grained datasets.The experimental results show that our method improves the classification accuracy of fine-grained images and achieves better classification performance.We also visualize the most discriminative region that be used to fine-grained classification,it shows that our model can effectively capture the subtle differences between subclasses,so that it can distinguish different subclasses effectively.(2)Fine-grained image features at different scales can provide different levels of information in the image for fine-grained image classification,and different levels of information can provide richer evidences for fine-grained image classification tasks.Therefore,we design a fine-grained image classification method based on multi-scale feature fusion.Specifically,we perform finer region localization in the discriminative regions to find more refined features that can be easily distinguished from other subclasses,such as the beak of a bird's head.Then,the weights are applied to fine-grained image features of different scales,so that our model focuses on the key regions with more discriminative feature in the image.Comparing with other classical fine-grained image classification methods,our method improves classification accuracy effectively.It shows that our method not only achieves discriminative region localization efficiently,but also utilizes different scales fine-grained features to rectify the final classification result of our model,so that it can further improve classification performance of fine-grained images.