The challenge in data compression is to transmit or store a minimum number of bits per sample that should be as close as possible to the number of bits defined by the actual average information content. The latter can be approximated by the contextual entropy. The dilemma is generally that the optimal estimation of contextual symbol probabilities is a kind of artificial-intelligence problem and requires enormous computational efforts. Therefore, practical compression methods typically utilise some pre-knowledge about the data. In image compression, this concerns either autocorrelation properties of natural images, such as photographs, or assumptions about repeating patterns in synthetic data that often can be observed in screen content. This paper presents a novel technique for the estimation of the contextual symbol probabilities with moderate computational complexity. The probability distribution is derived for each single pixel based on a soft context formation and is fed into a full-adaptive arithmetic coder. Applied to synthetic images and images with mixed content up to 8000 colors, the proposed scheme shows bit-savings of about 20% compared to the compression with the HEVC reference software (HM-16.7+SCM-6.0), and it also can compete with methods based on context mixing.