Research On Imbalanced Fine-grained Image Recognition Method Based On Deep Learning

With the development of artificial intelligence technology,people use this technology to realize the identification of different things in daily life.However,it is often necessary to distinguish the sub-categories of the target in daily life.Due to the lack of understanding of expert knowledge,people cannot accurately identify the target subclass,and it is difficult to accurately identify the target subclass only through traditional image recognition methods.How to improve the recognition method that can accurately recognition the sub-categories of target has gradually attracted the attention of computer vision field and become an important research direction in this field.Traditional image recognition methods aims to identify “cats” and “dogs”,which are easily distinguishable in terms of morphology and appearance,compared with the traditional image recognition,the difficulty of identifying sub-categories(that is,fine-grained image recognition)is the sub-categories images are more refined,and the characteristics of large intra-class differences and small inter-class differences will also make the recognition task more challenging.The research on fine-grained datasets has prompted different forms of fine-grained image recognition methods to be proposed,these methods improve the model recognition effect on sub-category from different aspects.However,the research of fine-grained datasets and their recognition methods found that there is an imbalance,which will affect the model’s recognition performance on fine-grained datasets.This paper takes the grape dataset and public dataset as the research objects,and based on the knowledge of deep learning,the imbalance of fine-grained datasets and their recognition methods is analyzed and solved from the perspectives of dataset category imbalance and detection method imbalance,respectively.The main research contents and results are as follows:(1)Constructed the grape long-tailed dataset(Vitis-15-LTs): The original grape dataset Vitis-15 contains the total of 6389 images from 15 grape varieties,this paper constructs the grape long-tailed dataset Vitis-15-LTs based on Vitis-15.By reviewing the shooting environment and grape planting characteristics of the grape dataset Vitis-15,it is found that the cultivation range of different grapes varieties is different due to factors such as sugar content,disease resistance,transportation potential and cultivation difficulty,which resulted in the long-tailed distribution of grape dataset collected from nature.To facilitate the simulation of the long-tailed distribution problem in grape dataset,the original grape dataset Vitis-15 was first augmented by the data augmentation method to make it a balanced dataset,and stratified sampling was used to divide the training set and the test set.Then the ratio of the training samples number for the head class and the tail class is defined as an imbalance factor,which is used to describe the severity of the long-tailed problem.Final,the long-tailed dataset Vitis-15-LTs with imbalance factors of10,50 and 100 were constructed.(2)Multi-expert distribution-aware network based long-tailed image recognition of grape: In order to solve the problem of long-tailed distribution in grape dataset,this paper adopts multi-expert distribution-aware network and compares its performance with typical long-tailed image recognition methods.By applying bias and variance analysis to typical long-tailed image recognition methods,it is found that these methods generally reduce tail bias by increasing model variance,but the gap between head and tail bias is still large,so the typical long-tailed image recognition methods were not the biggest extent reduce minimize the impact of long-tailed distributions on model performance.In order to ensure that the tail bias is reduced without increasing the variance of the model,and the gap between head and tail bias is minimized,the multi-expert distribution-aware network is used on the grape long-tailed datasets Vitis-15-LTs,which mainly includes multi-expert sharing modules and distribution-aware diversity loss.First,the multi-expert architecture is used to reduce the model variance of all classes,and then the expert sharing mode is used to reduce the problem of high memory consumption and high computational complexity caused by multi-expert network.Finally,the distribution-aware diversity loss will reduce the tail bias,and thus the model recognition performance for long-tailed dataset can be improved to the greatest extent.The experimental results on the grape long-tailed datasets Vitis-15-LTs show that the adopted method achieves lower false positive rate and more stable accuracy,and it achieves 96.83%,93.67% and 89.6% on the grape long-tailed datasets with three imbalance factors,respectively.(3)Balance based fragmentation learning for fine-grained image classification:The research on the strongly supervised fine-grained image recognition method proposed on the public fine-grained dataset found that the detection methods used in the‘localization-classification sub-network’ method are imbalance.In order to improve the detector performance limited by imbalance and improve the model’s recognition performance for fine-grained datasets,this paper proposes balance based fragmentation learning(BFL)method and compares the proposed method performance with fine-grained image recognition methods on multiple public fine-grained datasets.The balance based fragmentation learning method adopts three balancing strategies in the detection stage to improve the detection performance limited by the imbalance of samples,features and objects.In the classification stage,the image fragmentation mechanism is first used to change the global structure of the balanced detection image,and then the image blocks are randomly arranged to ensure that the network can find discriminative local regions in the image,and finally a more robust loss function is used to improve the model recognition performance on fine-grained datasets.Experimental results on three different categories of public fine-grained datasets CUB-200-2011,Stanford Car and FGVC Aircraft show that the proposed method achieves better recognition results than current fine-grained image recognition methods,and the test accuracy reached 89.1%,95% and 93.2%,respectively.