Work

SoftBinReduce: data reduction for color quantization through soft binning

Year: 2025

Type: article

Abstract: In this paper, we propose SoftBinReduce, a data reduction method for Color Quantization. Our approach consists of four main steps: (1) identifying representatives in each channel, (2) generating three-dimensional bins, (3) distributing pixel values using a soft binning technique, and (4) repositioning the resulting bins. Additionally, we introduce an adapted Sort-Means algorithm for the Pixel Mapping phase, providing a strong initial guess. We conduct an extensive experimental evaluation primarily using the CQ100 dataset, but also the older Kodak and USC-SIPI datasets. We compare our method to two well-known data reduction techniques from the literature: pseudo-random and quasi-random sampling. The results demonstrate that our method outperforms both in terms of NMSE error versus the achieved speedup, particularly when the data is significantly reduced and with a higher number of colors in the generated palette. Moreover, an ablation study shows that all components of our method are necessary for producing high-quality results.

Source: Multimedia Systems

DOI: 10.1007/s00530-025-01755-z

Authors: Guillem Rodríguez-Corominas, Maria J. Blesa, Christian Blum

Institutions: Consejo Superior de Investigaciones Científicas, Artificial Intelligence Research Institute, Universitat Politècnica de Catalunya

Language: English

Topic: Advanced Data Compression Techniques

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Estimated APC paid: $3,490