Local Contrast Criterion for Verifiable Image Enhancement Network: Layered Difference Representation Loss

Deep neural networks (DNN) have made significant improvements in image processing, particularly in media forensic investigations. However, the resulting images or frames from DNN-based algorithms are typically not admissible as evidence because these algorithms do not precisely verify the internal processes from input to output. This study proposes an efficient local contrast enhancement criterion for a layered difference representation (LDR) loss, a verifiable image enhancement network. The LDR originally derives a transformation function based on neighboring pixel value differences. However, appending an additional constraint, such as image similarity to the ground truth, is challenging. To address this, we utilize DNNs and introduce a novel criterion, LDR loss, for image enhancement. The LDR loss aims to increase neighboring pixel differences, whereas the image loss ensures similarity with the ground truth, thus enhancing both global and local contrasts of the image. Experimental results demonstrate that the proposed algorithm outperforms conventional image enhancement algorithms.