A retina-inspired neural model for single image haze removal
Yun Zhang, Qian Sun, Keyu Du, Hong Qu · 2023
The mammalian retina, a complex and highly specialized structure, serves a crucial role in transforming light into meaningful visual information. Drawing inspiration from the hierarchical processing of information in the retina’s neurons, we propose a multilayer retina-based convolutional neural model and apply this model to image dehazing. This model simulates the progression of information and functions from photoreceptor cells to bipolar cells in the retina. In the photoreceptor layer, we simulate cone neurons for color sensitivity and rod cells for brightness. Horizontal cells extract features from photoreceptor neurons and inhibit bipolar cells. Bipolar cells include on and off types, responding to increased and decreased light intensity, respectively. The amacrine cell layer excites on-bipolar cells while inhibiting off-bipolar cells. The ganglion cell layer simulates color-opponent mechanisms and disinhibitory effects. Finally, the combination of brightness and color information, along with spikes encoding, produces a haze-free image. This approach effectively reduces haze without estimating atmospheric light sources. The model’s validity is verified on two image de-hazing datasets: SOTS and D-Hazy, which encompass varying degrees of haze. Experimental results demonstrate that our proposed model has a comparable performance by comparing with the classical methods in de-hazing tasks.