Low-Level Salient Region Detection Using Multi-Scale Filtering

Gökhan Yıldırım, Sabine SüsstrunkInfoscience 2013

Abstract

Various computer vision tasks require a summary of visually important regions in an image. Thus, developing simple yet accurate salient region detection algorithms has become an important research topic. The currently best performing state-of-the-art saliency detection algorithms incorporate image segmentation for abstraction. However, errors introduced in this step of the algorithms are transferred to the final saliency map estimation. In order to avoid this problem, we propose a simple low-level salient region detection algorithm that uses multi-scale filters. We consider each possible combination of filtered image pairs as weak saliency maps and combine them according to their adaptively computed compactness and center prior. Our filter-based method successfully eliminates the texture in the background and gives relatively uniform salient regions for multi-colored objects. In addition, the combination of several multi-scale filters produces a full-resolution saliency output, which preserves object boundaries. We show that our algorithm outperforms the most recent state-of-the-art methods on a database of 1000 images with pixel-precision ground truths.

Introduction

Saliency detection—the process of identifying the most eye-catching regions in an image—is a foundational task in computer vision. Traditionally, top-performing algorithms have relied on image segmentation as a pre-processing step to abstract images into manageable regions. However, this "segmentation crutch" introduces significant risks: if the segmentation fails (e.g., over-segmenting texture or missing fine details), these errors propagate to the final saliency map. Furthermore, segment-based comparisons often produce "patchy" results for multi-colored objects.

This paper introduces a robust saliency detection algorithm that replaces traditional segmentation with a multi-scale filtering approach. By generating "weak" saliency maps from Gaussian filter pairs and weighting them with adaptive modules, the proposed method achieves state-of-the-art accuracy and superior boundary preservation.

Methodology: Multi-Scale Filtering Framework

The proposed framework operates in the CIELab color space and avoids segmentation entirely. The pipeline is built on three main pillars:

1. Multi-Scale Filtering

Instead of segmenting, the algorithm applies a series of 2D Gaussian filters of increasing sizes ($2^i + 1$). By calculating the pixel-wise square of differences between every possible pair of these filtered images, the system generates a set of weak saliency maps. Each map captures a specific band-pass spatial frequency interval. Small-scale differences act as edge detectors, while large-scale differences detect the object as a "blob," ensuring the interior is uniformly highlighted.

Weak saliency maps generated by various filter combinations.

2. Adaptive Compactness Measure ($K$)

To filter out background noise, the authors calculate the spatial distribution of salient pixels in each weak map. If "salient" pixels are widely dispersed (characteristic of background texture), the map is assigned a low compactness weight and suppressed.

3. Adaptive Center Prior ($P$)

Recognizing that salient objects are not always perfectly centered, the algorithm uses the center of mass from each weak map to adaptively shift a Gaussian prior mask. This allows the model to "follow" the object even if it is off-center, rather than blindly favoring the middle of the frame.

Experiments and Results

The algorithm was evaluated on the MSRA-1000 dataset, a benchmark of 1000 images with pixel-precision ground truth masks.

Quantitative Performance

The proposed method achieved an AUC (Area Under the Curve) of 0.890, outperforming the next best method (LR at 0.858) and significantly improving upon the filter-based baseline (FT at 0.629). Precision and recall values, balanced using Otsu’s adaptive thresholding, also showed superior performance.

Qualitative Comparison

As shown in Figure 1, the method produces cleaner suppression of textured backgrounds (like grass) and more uniform saliency for multi-colored objects compared to methods like RC, SF, and LR. Because it uses filters rather than segments, the output is full-resolution, preserving fine structures like the legs of the horse.

Comparison of the proposed method against state-of-the-art saliency detection algorithms.

Discussion and Conclusion

By moving away from the "segmentation crutch," this approach avoids blocky artifacts and recovers fine structures that segmentation often misses. The synergy between different filter scales allows for both boundary preservation and uniform object highlighting.

While the results are impressive, the current fusion of maps relies on fixed heuristic adjustment factors. Future work could involve a machine-learning approach to learn optimal fusion weights. Additionally, as a low-level spatial cue model, it does not yet incorporate high-level semantic context. Expanding the approach to larger datasets and incorporating semantic features remains a promising direction for future research.

Summary updated

Abstract

Various computer visions tasks require a summary of visually important regions in an image. Thus, developing simple yet accurate salient region detection algorithms has become an important research topic. The currently best performing state-of-the-art saliency detection algorithms incorporate image segmentation for abstraction. However, errors introduced in this step of the algorithms are transferred to the final saliency map estimation. In order to avoid this problem, we propose a simple low-level salient region detection algorithm that uses multi-scale filters. We consider each possible combination of filtered image pairs as weak saliency maps and combine them according to their adaptively computed compactness and center prior. Our filterbased method successfully eliminates the texture in the background and gives relatively uniform salient regions for multi-colored objects. In addition, the combination of several multi-scale filters produces a full-resolution saliency output, which preserves object boundaries. We show that our algorithm outperforms the most recent state-of-the-art methods on a database of 1000 images with pixel-precision ground truths.
BibTeX
@inproceedings{yldrm2013low,
  author = {Gökhan Yıldırım and Sabine Süsstrunk},
  title = {{Low-Level Salient Region Detection Using Multi-Scale Filtering}},
  booktitle = {Infoscience},
  year = {2013},
}
Figure