Illumination-Adaptive Edge Enhancement Using a Hybrid DSP–DL U-Net

Abstract

This paper presents a hybrid DSP–DL system for edge enhancement in images affected by severely degraded illumination, including under- and overexposed scenes. The proposed approach integrates deterministic gradientbased operators with an illumination-aware convolutional architecture to overcome the limitations of both classical edge extraction methods and learningbased detectors under non-uniform illumination conditions. A modified U-Net architecture is used to process a five-channel input tensor composed of histogramadjusted luminance, multiscale Sobel gradients (3×3 and 5×5), a map of global mean intensity, and a histogram-derived scene flag. Illumination adaptivity is introduced through a dynamic branching mechanism at the network input, to enable early modulation of feature extraction based on the luminance characteristics of the scene. A spatial attention module is introduced to support coherent edge reconstruction during decoding, while training of the network is guided by a dual loss formulation in which luminance consistency is balanced with gradient preservation through a learnable weighting parameter. An adaptive refinement stage further improves the edge continuity using luminance statistics and multiscale gradient cues. Quantitative evaluation is performed using five noreference (NR) edge and structure metrics, supported by statistical validation. Within the NR evaluation framework adopted here for illumination-degraded imagery, the proposed system demonstrates improved structural consistency relative to classical Sobel operators, and shows more stable behavior based on the NR metric when compared with representative pre-trained learning-based edge detectors under the same protocol and test conditions