DAE-Net: Diagnosis-Driven Adaptive Enhancement for Underwater Visual Big Data

Abstract

Marine engineering, ecological monitoring, and offshore industry now produce terabyte-scale streams of underwater imagery whose visual quality is degraded by three physically distinct mechanisms: wavelength-selective absorption (colour cast), low ambient irradiance (low light), and forward / backward scattering of suspended particulates (blur). These regimes co-occur in varying proportions and a single, uniform enhancement pipeline inevitably over-corrects one and under-corrects another. We present DAE-Net, a lightweight diagnosis-driven framework that (i) quantifies the three degradation regimes from joint RGB+CIELAB statistics into normalised scores $(\hat C,\hat L,\hat B) \in {[0, 1]^3}$, (ii) routes the input through three specialised enhancement branches—channel-rebalanced colour correction, multi-scale Retinex, and Wiener-filter deblurring—and (iii) recomposes their outputs with a diagnosis-driven softmax fusion whose weights are proportional to the diagnosed severity. On a 50-image synthetic densification derived from 12 public benchmark images, DAE-Net attains 44.6 UIQM and 15.2 UCIQE. Under the same protocol it improves PSNR and UIQM over HE, CLAHE, and the underwater dark-channel-prior baseline while running 2.6× faster than the latter; its UCIQE remains below CLAHE. Published UIEB-T90 scores for learning-based methods are reported only as cross-dataset context. The framework is interpretable, training-free, and branch-parallelisable, making it a lightweight candidate for low-rate underwater visual-data streams.