Abstract:In this paper, we propose a GraphBit method to learn unsupervised deep binary descriptors for efficient image representation. Conventional binary representation learning methods directly quantize each element according to the threshold without considering the quantization ambiguousness. The elements near the boundary dubbed as “ambiguous bits” fail to collect effective information for reliable binarization and are sensitive to noise that causes reversed bits. We argue that there are implicit inner relationships among bits in binary descriptors called bitwise interaction, where the related bits can provide extra instruction as prior knowledge for ambiguousness reduction. Specifically, we design a deep reinforcement learning model to learn the structure of the graph for bitwise interaction mining, and the uncertainty of binary codes is reduced by maximizing the mutual information with input and related bits. Consequently, the ambiguous bits receive additional instruction from the graph for reliable binarization. Moreover, we further present a differentiable search method (GraphBit+) that mines the bitwise interaction in continuous space, so that the heavy search cost caused by the training difficulties in reinforcement learning is significantly reduced. Since the GraphBit and GraphBit+ methods learn fixed bitwise interaction which is suboptimal for various input, the inaccurate instruction from the fixed bitwise interaction cannot effectively decrease the ambiguousness of binary descriptors. To address this, we further propose the unsupervised binary descriptor learning method via dynamic bitwise interaction mining (D-GraphBit), where a graph convolutional network called GraphMiner reasons the optimal bitwise interaction for each input sample. Extensive experimental results on the CIFAR-10, NUS-WIDE, ImageNet-100, Brown and HPatches datasets demonstrate the efficiency and effectiveness of the proposed GraphBit, GraphBit+ and D-GraphBit.

Unsupervised Deep Learning of Compact Binary Descriptors

Learning Compact Binary Descriptors with Unsupervised Deep Neural Networks

Deep Unsupervised Binary Descriptor Learning Through Locality Consistency and Self Distinctiveness

On Aggregation of Unsupervised Deep Binary Descriptor with Weak Bits

CDbin: Compact Discriminative Binary Descriptor Learned With Efficient Neural Network

Deep Hashing for Compact Binary Codes Learning

Learning Deep Binary Descriptor With Multi-Quantization

Learning Deep Binary Descriptors Via Bitwise Interaction Mining

DLBD: A Self-Supervised Direct-Learned Binary Descriptor

Large-Margin Learning of Compact Binary Image Encodings

SUBIC: A supervised, structured binary code for image search

Learning Compact Binary Face Descriptor for Face Recognition.

Tiny Descriptors for Image Retrieval with Unsupervised Triplet Hashing

Deep Hashing for Scalable Image Search.

Learning Transformation-Invariant Local Descriptors with Low-Coupling Binary Codes

A Dictionary Learning Based Unsupervised Neural Network for Single Image Compressed Sensing

Deep Supervised Hashing for Fast Image Retrieval

Two-Stage Unsupervised Deep Hashing For Image Retrieval

Image retrieval using compact deep semantic correlation descriptors

Towards compact visual descriptor via deep fisher network with binary embedding

Fast Image Search with Pixel-Based Deep Learning Framework Via Efficient Compact Binary Code and Addictive Latent Layer