Abstract
Pre-segmented pixels can reduce the difficulty of segmentation and promote the segmentation performance. This paper proposes a novel segmentation method based on merging texture superpixels by computing inner similarity. Firstly, we design a set of Gabor filters to compute the amplitude responses of original image and compute the texture map by a salience model. Secondly, we employ the simple clustering to extract superpixles by affinity of color, coordinates and texture map. Then, we design a normalized histograms descriptor for superpixels integrated color and texture information of inner pixels. To obtain the final segmentation result, all adjacent superpixels are merged by the homogeneity comparison of normalized color-texture features until the stop criteria is satisfied. The experiments are conducted on natural scene images and synthesis texture images demonstrate that the proposed segmentation algorithm can achieve ideal segmentation on complex texture regions.
Introduction
Image segmentation—the process of partitioning an image into meaningful regions—is a cornerstone of computer vision. However, traditional methods often struggle with complex textures due to two primary challenges: the Scale Gap, where individual pixels lack sufficient context to define a texture pattern, and the Homogeneity Trap, where distinct textures share similar average color profiles (e.g., a smooth green lawn vs. green leafy bushes).
This paper addresses these issues by introducing a two-stage framework that moves beyond simple color-based clustering. By integrating multi-scale texture features into both the initial superpixel generation and the subsequent region merging, the proposed method achieves higher boundary accuracy and regional consistency in complex natural scenes.
Methodology
The proposed pipeline is divided into two main phases: texture-aware superpixel extraction and region merging via compositive histograms.
1. Texture-Aware Superpixel Extraction
Unlike standard SLIC (Simple Linear Iterative Clustering) which relies solely on color and spatial coordinates, this method incorporates Texture Saliency:
- Gabor Filtering: A bank of Gabor filters (4 scales, 6 orientations) captures multi-directional texture responses.
- Dimensionality Reduction: High-dimensional Gabor features are compressed into a 1D texture map using a saliency model to avoid computational overhead.
- 6D Clustering: Pixels are clustered into superpixels based on a 6D feature vector: $[L, a, b, x, y, \text{TextureMap}]$. This ensures that superpixel boundaries respect texture transitions.
2. Region Merging via Compositive Histograms
Once the image is over-segmented, adjacent superpixels are merged based on similarity:
- Compositive Histograms: Each superpixel is described by a histogram combining HSV color and spatial distribution.
- Spatial Layout (CCV & LWR): To distinguish textures with identical colors, the method uses a Color Coherence Vector (CCV) and a Length-to-Width Ratio (LWR) for color clusters. This captures the geometric arrangement of colors within the superpixel.
- Adaptive Thresholding: Merging is controlled by an adaptive threshold ($J_{threshold}$) derived from the standard deviation of histogram features across the entire image, allowing the algorithm to adapt to varying image complexity.
Results and Discussion
The method was evaluated using the Berkeley Segmentation Dataset (BSD) and MIT Vistex textures, comparing it against established algorithms like Mean-Shift (MS), JSEG, and Unsupervised Graph-cut (UGC).
Quantitative Performance
On synthetic MIT Vistex images, the proposed method achieved a Probabilistic Rand Index (PRI) of 0.965 and an F-measure of 0.972, significantly outperforming JSEG (0.802 PRI / 0.692 F-measure). It also achieved the lowest Global Consistency Error (GCE) at 0.135, indicating high alignment with human perception.
Qualitative Analysis
Visual results on challenging subjects like zebras and leopards demonstrate the algorithm's ability to group complex patterns as single entities while maintaining sharp boundaries against backgrounds. As shown in the comparison below, the proposed method produces superpixels that better adhere to the intricate boundaries of texture-rich regions compared to SLIC or Tubor methods.
Conclusion
By treating texture as a fundamental feature alongside color and spatial proximity, this framework effectively bridges the gap between pixel-level data and regional patterns. While the merging stage introduces higher computational complexity (averaging 16 seconds per image on standard hardware), the gains in segmentation accuracy for complex natural scenes are substantial. Future work will focus on automatically determining the optimal initial number of superpixels and incorporating higher-level semantic features.
Summary updated