Adaptive Interval Type-2 Fuzzy Filter: An AI Agent for Handling Uncertainties to Preserve Image Naturalness

Artificial intelligence (AI) offers fuzzy set theory (FST) as one of the popular AI agents and decision making tools for digital image processing to increase the robustness of vision-based applications. FST is capable to handle uncertainties while enhancing image quality and preserving its naturalness. This article proposes a novel Adaptive Interval Type-2 Fuzzy Filter (AIT2FF) as an AI agent for preserving the naturalness of nonuniformly illuminated images. The proposed AIT2FF estimates coarse illumination of the input image, which is further processed to obtain the reflectance and refined coarse illumination for the composition of enhanced image. The estimated coarse illumination preserves the naturalness by eliminating uncertainties due to grayness ambiguities in homogeneous regions and spatial ambiguities at edges. The effectiveness of the proposed filter has been presented quantitatively in terms of lightness order error (LOE) and naturalness image quality evaluator (NIQE) score on images from the high dynamic range dataset and images captured with commercial digital cameras. The qualitative comparison visualizes that the enhanced images obtained using the proposed filter maintains visual realism and naturalness. The applications of the proposed filter have also been presented for image dehazing, vehicle tracking, and gradients estimation. Moreover, the detection results for dehazed images have been compared with state-of-the-art methods on real-world task driven testing set from REalistic Single Image DEhazing-
$\beta$
dataset. The comparisons demonstrate that the enhanced images obtained using the proposed AIT2FF approach are better and outperform others in terms of LOE and NIQE measures.

Impact Statement—The objective of vision-based applications is to automate the human visual perception. This requires a preprocessed input image with fine details. The image details to achieve human vision suffer from uncertainties in homogeneous regions and at edges. This result in an unnatural preprocessed image and degrades the performance. For example: Noisy images are not good to train deep learning models. AI offers one of the best tools, i.e., fuzzy set theory, which is capable to handle uncertainties. This article proposes an adaptive interval Type-2 fuzzy filter for elimination of uncertainties. The proposed filter enhances the image details and maintains its natural appearance. The elimination of uncertainties using proposed filter drops the error value from 2.573 to 1.929 for nonuniformly illuminated images. The proposed approach is able to boost the performance of various vision-based applications like object detection, vehicle tracking, security, medical diagnostics, etc.