Low-Level Features for Image Retrieval Based on Extraction of Directional Binary Patterns and Its Oriented Gradients Histogram

In this paper, we present a novel approach for image retrieval based on extraction of low level features using techniques such as Directional Binary Code, Haar Wavelet transform and Histogram of Oriented Gradients. The DBC texture descriptor captures the spatial relationship between any pair of neighbourhood pixels in a local region along a given direction, while Local Binary Patterns descriptor considers the relationship between a given pixel and its surrounding neighbours. Therefore, DBC captures more spatial information than LBP and its variants, also it can extract more edge information than LBP. Hence, we employ DBC technique in order to extract grey level texture feature from each RGB channels individually and computed texture maps are further combined which represents colour texture features of an image. Then, we decomposed the extracted colour texture map and original image using Haar wavelet transform. Finally, we encode the shape and local features of wavelet transformed images using Histogram of Oriented Gradients for content based image retrieval. The performance of proposed method is compared with existing methods on two databases such as Wang’s corel image and Caltech 256. The evaluation results show that our approach outperforms the existing methods for image retrieval.

💡 Research Summary

The paper proposes a multi‑stage low‑level feature extraction pipeline for content‑based image retrieval (CBIR) that combines color‑texture, frequency, and shape information. First, each of the three RGB channels is processed independently with Directional Binary Code (DBC). Unlike the classic Local Binary Pattern (LBP), which compares a central pixel with its eight neighbours, DBC compares two neighbouring pixels along a predefined direction (e.g., 0°, 45°, 90°, 135°) and encodes the result as a binary value. By repeating this for several directions, DBC captures richer spatial relationships and edge information. The three DBC maps are merged to form a color‑texture representation that simultaneously reflects chromatic and textural cues.

Second, both the original image and the derived color‑texture map are decomposed using a two‑level Haar wavelet transform. Haar wavelets separate the image into low‑frequency (approximation) sub‑bands that preserve overall shape and color, and high‑frequency (detail) sub‑bands that contain edges and fine texture. This multi‑resolution decomposition allows the subsequent descriptor to exploit information at different scales.

Third, Histogram of Oriented Gradients (HOG) is computed on each wavelet sub‑band. HOG divides a sub‑band into cells, calculates gradient orientation histograms, and normalizes them across blocks, thereby providing a robust representation of local shape and edge distribution. Applying HOG to both low‑ and high‑frequency components yields a concatenated feature vector that encodes shape information across scales, while the preceding DBC step already supplies a detailed color‑texture signature.

The authors evaluate the method on two public datasets: Wang’s Corel image collection and Caltech‑256. They compare against baseline approaches such as LBP‑HOG, Local Ternary Pattern (LTP)‑HOG, and a recent deep‑learning baseline that extracts features from a pre‑trained VGG‑19 network. Retrieval performance is measured using precision, recall, and mean average precision (mAP). The proposed pipeline consistently outperforms the baselines, achieving roughly 3–5 percentage‑point gains in mAP. The improvement is most pronounced on images with rich color variation and complex backgrounds, where DBC’s directional texture encoding proves advantageous.

Strengths of the work include: (1) DBC’s ability to capture directional edge and corner information beyond what LBP offers; (2) the use of Haar wavelets to separate global shape from fine‑grained texture, facilitating complementary feature fusion; (3) multi‑scale HOG descriptors that robustly encode shape across resolutions.

However, the approach has notable limitations. DBC’s computational cost grows linearly with the number of directions; the authors employ eight directions, resulting in roughly 2–3× the runtime of LBP. Haar wavelets, while computationally cheap, can overly smooth very thin structures and are sensitive to high‑frequency noise. Moreover, the pipeline is entirely handcrafted and lacks learnable parameters, limiting its adaptability to new domains compared with end‑to‑end deep models.

Future work suggested by the authors includes learning optimal direction sets in a data‑driven manner to reduce DBC overhead, experimenting with more sophisticated wavelet families (e.g., Daubechies, Symlet) to improve high‑frequency fidelity, and integrating CNN‑based local descriptors in place of or alongside HOG to boost discriminative power. Such extensions could preserve the interpretability and low computational footprint of the current method while narrowing the performance gap with modern deep‑learning‑based CBIR systems.