Image Retrival

Introduction

Image retrieval, including feature retrieval and segmantic retrieval, has been widely applied and made impressive progress in many related applications, such as face recognition and image searching. However, there still remain some challenging problems. For feature retrieval in large-scale dataset, high-efficiency, robust and general algorithms are required to deal with severe noises and complex situations. For segmantic retrieval in practical situations, it always suffers from semantic diversity, sparseness, uneven distribution and lack of labels, known as “Few-shot” or “Zero-shot”. Facing these challenges, we proposed several frameworks, including  Sparse Hashing with Optimized Anchor Embedding, Robust and General VQ Framework (RGVQ), Sample-transfer Zero-shot Learning and Deep Confidence Network for Robust Image Classification (DECODE). These techniques above have shown great performance in real-world applications.

Highlights

Sparse Hashing with Optimized Anchor Embedding

• Optimization on the integration of the anchors

• Pushing the anchors far from the axis while preserving their ralative positions

• Formulated as an orthogonality constrained maximazation problem

• Performance & analysis on five benchmark datasets

Guo, Y., Ding, G., Liu, L., Han, J., & Shao, L. (2017). Learning to hash with optimized anchor embedding for scalable retrieval. IEEE Transactions on Image Processing, 26(3), 1344-1354.

Robust and General Vector Quantization

• ℓ_p,q-norm loss function

• ℓ_p-norm to conduct similarity search

• q-th order loss to enhance robustness

• Novel optimization scheme to minimize non-smooth and non-convex ℓ_p,q-norm function

Guo, Y., Ding, G., & Han, J. (2018). Robust quantization for general similarity search. IEEE Transactions on Image Processing, 27(2), 949-963.

Zero-shot Learning with Transferred Samples

• One-step recognition framework to perform recognition in original feature space

• Sample transfer and pseudo labeling based on transferbilityand diversity

• Modified SVM to recognize unreliable positive samples

Guo, Y., Ding, G., Han, J., & Gao, Y. (2017). Zero-shot learning with transferred samples. IEEE Transactions on Image Processing, 26(7), 3277-3290.

DECODE: Deep Confidence Network for Robust Image Classification

• Determine the confidence that a sample is mislabeled

• Different weight to different samples to make the model pay less attention to low confidence data

• Easily combined with exsting architectures

Ding, G., Guo, Y., Chen, K., Chu, C., Han, J., & Dai, Q. (2019). DECODE: Deep Confidence Network for Robust Image Classification. IEEE Transactions on Image Processing.