Deep Learning Based Fine-Grained Species Identification
- Publication Type:
- Thesis
- Issue Date:
- 2021
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Fine-Grained Visual Categorization (FGVC) is a challenging research topic in computer vision. It deals with the classification of visual data at a subordinate level. This thesis investigates four categories of FGVC methods based on deep learning, including general convolutional neural networks, object part localization methods, approaches using CNN ensemble or higher-order feature encoding, and methods utilizing recurrent visual attention. Overall performance comparison has been conducted to analyse their advantages and disadvantages.
We proposed a new regression-based part detection structure and a novel part-based model, which increased the classification accuracy of PS-CNN from 76.4% to 82.4% on the CUB-200-2011 benchmark dataset.
Inspired by the second-order pooling, we proposed a highly interpretable method with a compressed structure to significantly reduce the computation complexity while improving the fine-grained categorization accuracy. The proposed model provides a supervised selection of the most discriminative second-order channels. With the proposed method, the computation and the feature dimension are linearly reduced to 4% of the original bilinear pooling. By applying matrix normalization and a Fisher-Recurrent-Attention structure, we achieved the best result among the VGG-16 based FGVC models.
Following the conception of attention crop and attention drop in the Fisher-Recurrent-Attention model, we proposed a forcing module to constrain the network to extract more diverse features for FGVC. The forcing module focuses more on confusion regions which are essential for the fine-grained classification. Experimental results show that the proposed forcing module can improve the attention and prediction of the network when an input image is panned or zoomed, and the double prediction performs better than the single prediction.
The existing FGVC methods often come with enormous amounts of computation and require large memory space. This makes these models inadequate for mobile applications. We proposed a Category Attention Transferring Convolutional Neural Network (CAT-CNN) to transfer the attention knowledge from a large-scale FGVC network to a small but efficient network to improve its presentation capability. Using the proposed model, we improved the classification accuracy of the efficient networks by up to 5.7% on the CUB-2011-200 dataset without increasing computation time or memory cost, which makes FGVC feasible on mobile devices.
We also conducted abundant studies to investigate the relationship between attention and classification accuracy of our proposed deep learning models, visualized and analysed the attentional activations of these models. We hope that our findings may inspire further research efforts to advance the FGVC for a wide range of real-world applications.
Please use this identifier to cite or link to this item: