Retina-Inspired and Physically Based Image Enhancement

Images and videos with good lightness and contrast are vital in several applications, where human experts make an important decision based on the imaging information, such as medical, security, and remote sensing applications.
The well-known image enhancement methods include spatial and frequency enhancement techniques such as linear transformation, gamma correction, contrast stretching, histogram equalization and homomorphic filtering. Those conventional techniques are easy to implement but do not recover the exact colour of the images; hence they have limited application areas. Conventional image/video enhancement methods have been widely used with their different advantages and drawbacks; since the last century, there has been increased interest in retina-inspired techniques, e.g., Retinex and Cellular Neural Networks (CNN) as they attempt to mimic the human retina. Despite considerable advances in computer vision techniques, the human eye and visual cortex by far supersede the performance of state-of-the-art algorithms.

This research aims to propose a retinal network computational model for image enhancement that mimics retinal layers, targeting the interconnectivity between the Bipolar receptive field and the Ganglion receptive field.

The research started by enhancing two state-of-the-art image enhancement methods through their integration with image formation models. In particular, physics-based features (e.g. Spectral Power Distribution of the dominant illuminate in the scene and the Surface Spectral Reflectance of the objects contained in the image are estimated and used as inputs for the enhanced methods). The results show that the proposed technique can adapt to scene variations such as a change in illumination, scene structure, camera position and shadowing. It gives superior performance over the original model.

The research has successfully proposed a novel Ganglion Receptive Field (GRF) computational model for image enhancement. Instead of considering only the interactions between each pixel and its surroundings within a single colour layer, the proposed framework introduces the interaction between different colour layers to mimic the retinal neural process; to better mimic the centre-surround retinal receptive field concept, different photoreceptors' outputs are combined. Additionally, this thesis proposed a new contrast enhancement method based on Weber's Law.

The objective evaluation shows the superiority of the proposed Ganglion Receptive Field (GRF) method over state-of-the-art methods. The contrast restored image generated by the GRF method achieved the highest performance in contrast enhancement and luminance restoration; however, it achieved less performance in structure preservation, which confirms the physiological studies that observe the same behaviour from the human visual system.