Abstract
In order to quickly retrieve interested images from mass image databases, this paper proposes a composite descriptor based on contour shape and spatial features. We first extract the contour using a Shape Matrix Fourier Descriptor (SMFD). Then, adjacent contour sampling points and the image center are used to form a triangle, which has a unique circumcircle. The radii of these circumcircles form a distance histogram. By adjusting weights to combine these two methods, we measure similarity using Euclidean distance. Experimental results demonstrate that this method is simple, fast, and invariant to translation, rotation, and scaling, achieving superior retrieval accuracy and recall rates compared to existing methods.
Introduction
Content-Based Image Retrieval (CBIR) remains a significant challenge due to the "semantic gap" between low-level visual features and high-level human perception. While statistical descriptors like standard Fourier or Wavelet transforms capture global contour distributions, they often neglect local spatial relationships. This paper addresses this by introducing a composite descriptor that integrates global statistical signatures with local geometric layouts.
Methodology
The proposed approach merges two distinct feature types into a unified similarity metric.
Shape Matrix Fourier Descriptor (SMFD)
The process begins with contour extraction via the Canny operator. The shape is then mapped onto a polar coordinate grid centered at the centroid, creating a "shape matrix" where rows represent concentric rings and columns represent rays. To achieve rotation invariance, this matrix is converted into a 1D vector and processed with a Fourier Transform. Noise is reduced by retaining only coefficients with a modulus higher than the average.
Circumcircle-Based Spatial Feature (HD)
To capture local geometry, the centroid and two adjacent contour sampling points form a triangle. The radius of the unique circumcircle for each triangle is calculated along the contour to build a Histogram of Distances (HD). This feature effectively encodes local curvature relative to the object's center.
Similarity Measurement
The final composite descriptor combines the Euclidean distances of the SMFD ($D_{SMFD}$) and the HD ($D_{HD}$) using a weighted sum: $$D_{total} = \omega_{SMFD} \cdot D_{SMFD} + \omega_{HD} \cdot D_{HD}$$ Optimal performance was found with weights of $\omega_{SMFD} = 0.6$ and $\omega_{HD} = 0.4$.
Experimental Results
The method was evaluated using the MPEG-7 benchmark dataset, comprising 70 categories with 20 variations each. The proposed method (M1) was compared against Shape Contexts with Dynamic Programming (M2) and standard contour Fourier transforms (M3).
As shown in the performance curves, M1 consistently outperforms M2 and M3 in both precision and recall across various categories such as "Apple," "Bat," and "Bird." For instance, in the "Apple" category, M1 achieved approximately 71% precision, significantly higher than M2 (~50%) and M3 (~60%). Furthermore, M1 is computationally more efficient than the dynamic programming-based M2, making it suitable for large-scale retrieval.
Conclusion
The composite descriptor provides a robust and efficient solution for shape-based image retrieval. By combining global contour statistics with local spatial geometry, it bridges the semantic gap more effectively than traditional methods. Future work will focus on optimizing sampling density and improving robustness in complex, non-silhouetted backgrounds.
Summary updated