Information-based scale saliency methods with wavelet sub-band energy density descriptors
Faculty of Health, Engineering and Science
School of Engineering / Centre for Communications and Electronics Research
Pixel-based scale saliency (PSS) work bases on information estimation of data content and structure in multiscale analysis; its theoretical aspects as well as practical implementation are discussed by Kadir et al . Scale Saliency framework  does not work only for pixels but other basis-projected descriptors as well. While wavelet atoms, localization in both time and frequency domain, are possible alternative descriptors, no theoretical analysis and practical solutions have been proposed yet. Our contribution is introducing a mathematical model of utilizing wavelet-based descriptors in a correspondent Wavelet-based Scale Saliency (WSS). It treats wavelet sub-band energy density of two popular discrete wavelet transform (DWT) and dual-tree complex wavelet transform (DTCWT) as basis descriptors instead of pixel-value descriptors for saliency map estimation. Then, ROC, AUC, and NSS quantitative analysis are comparing WSS against PSS as well as other state-of-the-art saliency methods ITT , SUN , SRS  on N. Bruce's database  with human eye-tracking data as ground-truth. Furthermore, qualitative results, different saliency maps, are analyzed case by case for their pros and cons; especially their short-comings in specific situation or insensible results for human perception.