Deep Binary Representation for Efficient Image Retrieval

Abstract

With the fast growing number of images uploaded every day, efficient content-based image retrieval becomes important. Hashing method, which means representing images in binary codes and using Hamming distance to judge similarity, is widely accepted for its advantage in storage and searching speed. A good binary representation method for images is the determining factor of image retrieval. In this paper, we propose a new deep hashing method for efficient image retrieval. We propose an algorithm to calculate the target hash code which indicates the relationship between images of different contents. Then the target hash code is fed to the deep network for training. Two variants of deep network, DBR and DBR-v3, are proposed for different size and scale of image database. After training, our deep network can produce hash codes with large Hamming distance for images of different contents. Experiments on standard image retrieval benchmarks show that our method outperforms other state-of-the-art methods including unsupervised, supervised, and deep hashing methods.

Deep hash for latent image retrieval

“…Different from DSH, SUBIC (Structured binary codes, SUBIC) [15] proposed by Jain et al introduced a block-SoftMax structure in the design process of loss function and verified the validity of the structure with many experiments. Lu et al [16] pre-calculated the hash codes of images by designed algorithm, and then fit them through deep neural networks. Compared with the method that uses category as ground truth, its hash code has more ability to tell differences with different images.…”

“…The hash-based image retrieval technology performs binary hash encoding based on the content of the images, and it uses the Hamming distance between the hash codes to measure the similarity of different images. Compared with traditional content-based image retrieval technology, the image feature adopted by hash-based image retrieval technology is binary hash coding, which has better storage and computational advantages [1]. Thus, it is possible to perform fast image retrieval within limited storage space.…”

“…It has the advantage that the obtained binary hash code is achieved directly from the original image and does not go through the image feature extraction process, so the error that may accused in the image feature extraction process can be reduced. In addition, lots of experiments [11][12][13][14][15][16] have also verified that the image hash retrieval method based on convolutional neural network has better retrieval performance than the traditional image hash retrieval method, and its retrieval performance is not very dependent on the hash code length.…”

Learning Hash Codes via Hamming Distance Targets

“…Alternative methods to force outputs close to ±1 exist, such as HashNet, which gradually sharpens sigmoid functions on the pre-binarized outputs. Another family of methods first learns a target hash code for each class, then minimizes distance between each embedding and its target hash code [21,15].…”

“…We highlight 5 comparator models: DBR-v3 [15], HashNet [2], Deep hashing network for efficient similarity retrieval (DHN) [23], Iterative Quantization (ITQ) [5], and LSH [4]. DBR-v3 learns by first choosing a target hash code for each class to maximize Hamming distance between other target hash codes, then minimizing distance between each image's embedding and target hash code.…”