Learning Deep Binary Descriptors via Bitwise Interaction Mining.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Conventional binary representation learning methods struggle with quantization ambiguity, leading to unreliable binarization and noise sensitivity.
• Implicit relationships among bits, termed bitwise interaction, can reduce this ambiguity by providing prior knowledge.

Purpose of the Study:

• To develop novel unsupervised deep binary descriptor learning methods for efficient image representation.
• To address the limitations of existing methods by effectively mining and utilizing bitwise interactions.

Main Methods:

• GraphBit: Utilizes deep reinforcement learning to mine bitwise interaction graphs, reducing binary code uncertainty.
• GraphBit+: Employs a differentiable search for continuous space mining, reducing computational cost.
• D-GraphBit: Introduces dynamic bitwise interaction mining using a graph convolutional network (GraphMiner) for input-specific optimization.

Main Results:

• The proposed methods (GraphBit, GraphBit+, D-GraphBit) demonstrate significant efficiency and effectiveness.
• Experiments on diverse datasets (CIFAR-10, NUS-WIDE, ImageNet-100, Brown, HPatches) validate the performance.
• D-GraphBit shows improved ability to decrease ambiguousness by dynamically mining interactions.

Conclusions:

• Unsupervised deep binary descriptor learning can be significantly enhanced by mining bitwise interactions.
• Dynamic mining of bitwise interactions offers superior performance over fixed interaction strategies.
• The proposed methods offer efficient and effective solutions for image representation and retrieval.