TPAMI: Frequency Domain Scale-Space Analysis for Visual Saliency Detection

This paper presents a visual saliency detection method based on scale-space analysis in the frequency domain, published in the top-tier journal TPAMI. The approach predicts human eye-attention regions by analyzing multi-scale frequency characteristics of images in the spectral domain. The core methodology involves first transforming the image into the frequency domain using Fourier transform, then analyzing frequency component variations across different scales. By comparing frequency response differences between scales, the algorithm effectively captures salient regions that contrast with their surroundings. Compared to traditional spatial domain methods, this frequency domain analysis better handles complex backgrounds and multi-scale objects. The key innovation lies in integrating scale-space theory with frequency domain analysis, eliminating the need for manual feature engineering in conventional approaches. The algorithm naturally implements multi-scale processing in the spectral domain, where saliency is automatically detected through frequency component variations. This method shows particular advantages in object detection within complex scenes, especially when handling texture-rich images or those with illumination variations. From an implementation perspective, the method typically involves: 1. Applying 2D Fourier transform to convert spatial image to frequency domain 2. Constructing scale-space representation using Gaussian filters in frequency domain 3. Calculating spectral residuals across different scales 4. Inverse transforming the salient frequency components back to spatial domain This research holds significant importance for computer vision, with the proposed framework applicable to object recognition, image segmentation, and visual compression. The frequency domain analysis provides a novel perspective for visual saliency detection, and this approach can be extended to other image analysis tasks.