Regularized Learning for Discriminative Problems in Image Recognition
Open Access
- Author:
- Li, Xuelu
- Graduate Program:
- Electrical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- June 14, 2021
- Committee Members:
- Puneet Singla, Outside Unit & Field Member
William Higgins, Major Field Member
Vishal Monga, Chair & Dissertation Advisor
Kultegin Aydin, Program Head/Chair
Jing Yang, Major Field Member - Keywords:
- Histopathological Image Classification
Fine-grained Image Classification
3-D Vessel Segmentation
Sparse Representation based Classification
Deep Learning
Shared Feature Learning - Abstract:
- Discriminative problems are centrally important in the computer vision area and have been applied to many different fields. The most common discriminative problems in computer vision include both image-level and pixel-level classification. Traditional discriminative problems first try to design special feature extractors to extract the representative discriminative information from the images. Then, a classification scheme is designed to predict the correct labels. Such classification techniques are restricted by their limitations such as the requirement of exhaustive feature engineering, and the extra burden brought by the separate training processes of the feature extractor and the classifier. Nowadays, learning based methods have gained more attention due to their flexibility in feature extraction and strong practical performance over traditional methods. Among them, sparse representation based classification and deep learning based techniques are the most effective learning based methods. In this proposal, we utilize these different learning based methods to solve open challenges in both the image and pixel-level classification. In particular, our methods are designed to exploit domain based knowledge hence are able to extract more discriminative information. The first part of this proposal is about the application of sparse representation based classification for histopathological image classification. The diversity of tissue structure in histopathological images makes feature extraction for classification a challenging task. We propose a novel Analysis-synthesis model Learning with Shared Features algorithm (ALSF) for classifying such images more effectively. In ALSF, a joint analysis and synthesis learning model is introduced to learn the classifier and the feature extractor at the same time. Unlike SRC, no explicit optimization is needed in the inference phase leading to much reduced computation. Crucially, we introduce the learning of a low rank shared dictionary and a shared analysis operator, which more accurately represents both similarities and differences in histopathological images from distinct classes. We also develop an extension of ALSF with a sparsity constraint, whose presence or absence facilitates a cost-performance trade-off. Experimental results demonstrate both the complexity and performance benefits of ALSF over state-of-the-art alternatives. Significance: Modeling shared features with appropriate quantitative constraints leads to significantly improved classification in histopathology. The second part of the proposal extends the idea of shared features between classes for the first time to deep learning. Given that images from distinct classes in fine-grained classification share significant features of interest, we present a new deep network architecture that explicitly models shared features and removes their effect to achieve enhanced classification results. Our modeling of shared features is based on a new group based learning wherein existing classes are divided into groups and multiple shared feature patterns are discovered (learned). We call this framework Group based deep Shared Feature Learning (GSFL) and the resulting learned network as GSFL-Net. Specifically, the proposed GSFL-Net develops a specially designed autoencoder that is constrained by a newly proposed Feature Expression Loss to decompose a set of features into their constituent shared and discriminative components. A key benefit of our specialized autoencoder is that it is versatile and can be combined with state-of-the-art fine-grained feature extraction models and trained together with them to improve their performance directly. Experiments on benchmark datasets show that GSFL-Net can enhance classification accuracy over the state of the art with a more interpretable architecture. In the final contribution of the thesis, we solve a more difficult pixel-level classification problem in image recognition --- 3-D vessel segmentation. Inspired by the observation that 3-D vessel structures project onto 2-D image slices with highly informative and unique edge profiles, we propose a novel, customized edge profile guided deep network for 3-D vessel segmentation. Our network architecture learns both edge profiles and segmentation maps jointly with the help of a shared encoder. 3-D context is mined in both the segmentation and edge prediction branches by employing long-short term memory (LSTM) modules. As a key contribution, we develop novel regularization terms that: a) capture the homogeneity of blood vessel images in the presence of biomarkers, and b) ensure robustness to domain-specific noise. The proposed method is referred to as Structural Prior Guided Deep 3-D Vessel Segmentation (SPG-3DVS). Experiments on benchmark datasets reveal that the proposed approach outperforms state-of-the-art techniques on standard measures such as dice overlap and mean Intersection-over-Union. Overall, this dissertation establishes through distinct examples that domain inspired and prior guided regularization in learning can help address challenges such as limited training and noise robustness. Such domain enriched regularization may be combined with alternatives such as transfer learning or incremental learning in future work.