Image retrieval method based on local projection andblock LBP feature

Bin Zou, Zhibin Pan, Sen HuJournal of Image and Graphics 2012 · DOI

Abstract

This paper presents a robust Content-Based Image Retrieval (CBIR) framework designed to overcome the limitations of traditional color histograms and pixel-level texture analysis. The proposed method integrates three core components: a projection-based index histogram for capturing spatial color distributions, a novel "Block-LBP" for noise-resilient texture extraction, and a saliency-weighted feature scheme to prioritize visually significant regions. Experimental results on the Corel dataset demonstrate a significant performance gain, with an average precision increase of 6.39% over standard indexing methods.

Introduction

Content-Based Image Retrieval (CBIR) remains a fundamental challenge in computer vision, aiming to retrieve images from large databases based on visual cues like color, texture, and shape. Traditional approaches often struggle with the "semantic gap"—the discrepancy between low-level pixel data and high-level human perception. Specifically, global color histograms lack spatial context, while standard Local Binary Patterns (LBP) are often overly sensitive to noise and computationally intensive. This research addresses these gaps by introducing localized projection features and a block-based texture descriptor, further refined by a visual saliency model.

Methodology

The authors propose a three-pronged approach to extract more meaningful "fingerprints" from images:

1. Local Projection and Vector Quantization (VQ)

To capture both color distribution and spatial relationships, the image is partitioned into $4 \times 4$ sub-blocks. Within each block, color intensities are projected along four primary directions: horizontal, vertical, and two diagonals. To ensure a compact representation, a Discrete Cosine Transform (DCT) is applied to these projection vectors, retaining only the most significant low-frequency coefficients. These coefficients are then compressed via Vector Quantization (using the LBG algorithm) to generate a compact "index histogram" that serves as a spatial-color fingerprint.

2. Block-based Local Binary Patterns (Block-LBP)

Standard LBP operates at the pixel level, making it susceptible to local fluctuations and noise. The proposed Block-LBP instead calculates texture features by comparing the average intensities of $2 \times 2$ pixel neighborhoods within a $3 \times 3$ grid of blocks. This approach provides two major benefits: it enhances robustness against image noise and drastically reduces computational complexity—estimated at approximately 1/16th the load of traditional LBP.

3. Saliency-Weighted Retrieval

Human vision naturally prioritizes foreground objects over background clutter. To mimic this, the framework employs the Itti saliency model to segment images into "salient" (likely foreground) and "non-salient" (background) regions. During the retrieval process, features extracted from salient regions are assigned higher weights ($\omega_1 > \omega_2$). This ensures that the similarity metric is dominated by the primary subject of the image, bridging the gap between raw data and visual intent.

Experimental Evaluation

The system was evaluated using 1,000 images from the Corel database, spanning 10 categories such as "Africa," "Dinosaurs," and "Horses." Each image served as a query to test the retrieval accuracy against the rest of the set.

Key Results: * Overall Accuracy: The proposed method achieved an average precision of 75.0%, compared to 68.0% for the baseline index histogram method—a 6.39% improvement. * Category Performance: The system performed exceptionally well in categories with distinct foreground subjects, such as "Dinosaur" (99.2% precision) and "Horse" (94.4%). * Robustness: Even in complex categories like "Beach," the inclusion of saliency weighting and robust Block-LBP features led to a precision increase from 37.6% to 47.0%.

Conclusion

By combining localized color projections with a noise-resilient Block-LBP and a saliency-driven weighting scheme, this framework provides a more effective and computationally efficient solution for CBIR. Future work will focus on optimizing the saliency model for real-time applications and exploring automated weight selection to further enhance retrieval performance across diverse image domains.

Summary updated

BibTeX
@article{zou2012image,
  author = {Bin Zou and Zhibin Pan and Sen Hu},
  title = {{Image retrieval method based on local projection andblock LBP feature}},
  journal = {Journal of Image and Graphics},
  year = {2012},
  doi = {10.11834/jig.20120609},
}
Figure