Research On Fine-grained Insects Image Classification Based On Weak Supervision

Fine-grained Image Classification is a very important problem in Computer Vision,and it can be applied to many professional and valuable scenarios.Even there have been a large number of deep learning applications of insect recognition,there is no research on the fine-grained image classification algorithms.Relying on the self-built fine-grained image classification dataset,this thesis designs some special deep learning tricks for the dataset,uses a series of weakly supervised algorithms on it and make improvement,and proposes a hierarchical multi-label constraint optimization fine-grained insect classification framework based on metric learning.The main work of this thesis is as follows:(1)Establish a fine-grained insect image dataset,which includes hard samples of fine-grained classification with small inter-class variance and large intra-class variance.Target to sovle the problem of the dataset,an online hue jitter data augment process with tag-control and a classification loss function with online-count weight are proposed.(2)Deeply analysis the common weak-supervision fine-grained classification algorithms.Combined with comparative experiments,this thesis trained a residual network which is suitable for this task and has good performance as the baseline network.Besides,the channel attention structures are inserted to different positions of the network to extract different levels of visual feature.And a single-stream bilinear pooling operation,which is tuned from the two-stream bilinear pooling network,is added to the network structure.And a dimensionality reduction method is used to reduce the parameter amount of the bilinear pooling operation.(3)The common fine-grained image classification algorithm only applies single finegrained tag as supervision information.In the field of insect identification,due to the multi-layer insect tags,especially the upper-level tags,are easy to get,this thesis establishs a three-level-tag fine-grained insect image dataset,and proposes a new evaluation metric to evaluate the classification performance of algorithm.Besides,a classification optimization framework with hierarchical multi-label as the supervision information and the loss function in the metric learning as the constraint is proposed.The framework can be used on each kind of fine-grained classification network.And this thesis studies the useful loss function and structure of the framework on the dataset of this thesis.(4)The improved weakly supervised fine-grained classification model and multilabel constrained optimization framework proposed in this paper performs much better than the Baseline model in this paper.The effectiveness of each structure is proved in the process of its proposed.