Computationally Efficient CNN Based Smart Dehaze Net for Haze Removal of Biomedical Images

Artificial Intelligence has made significant strides recently especially in the field of computer vision. It played a pivotal role in revolutionising the automation, reliability and robustness of the most modern biomedical applications like CT scan, PET scan, Fundoscopy etc. Due to the wide spread use of camera sensors for examining the internal organs like Liver, Gall bladder, Retina etc, haze free images are absolutely essential. Most of the images captured by the visual sensors have haze introduced due to various environmental factors like suspended particles in air which badly affects the fidelity of the images captured by camera sensors. Image dehazing is a process of removing the haze and thereby improving the overall image quality. Image dehazing finds extensive applications especially in the fields of surveillance systems, defence, air traffic control, traffic cameras, self-driven cars, under water imaging etc. In this paper a computationally efficient CNN architecture with eight convolution and three concatenation layers is proposed which removes the haze without much degradation in colour and contrast of the image. The Proposed model is trained on NYU2 dataset and provided a 166% hike in the PSNR when compared with the best of the available haze removal techniques. Also the BRISQUE value and entropy of the haze removed image went up by 80.95 %. and 100.78% respectively over the current state of the art haze removal techniques. Also the model shown 59% reduction in computational time.

I. Introduction

Undoubtedly clear vision is the most important criteria when it comes to low latency applications like self-driven cars, real time crowd management, traffic signaling etc. Recently computer vision systems [1] has made drastic evolution which is reinstated by their widespread application in various fields. These include surveillance [1], detection of objects in images [2], decomposition of images [3] to extract finer regions in an image, reduction of noise in images [4], analysis of crowd behaviour [5] etc. We often encounter image degradation due to foggy weather conditions. With the recent advancements in the field of camera imaging, the researchers are keen to develop new imaging technologies powered by AI. This is primarily due to the fact that in most of the image dehazing, crystal clear output images are not generated out of hazy images. Very often the dehazed image suffer from contrast degradation and colour distortion. The Camera lens captures the light that is directly reflected from the object as well as the scattered light due to the suspended particles in the medium i.e. air. [6]. Clear image visibility is crucial for achieving high precision in many artificial intelligence (AI) and computer vision applications, particularly in tasks like detection of objects and its recognition. The contemporary dehazing techniques can be broadly classified into two types 1) Earlier Prior based and Learning based methods. The first method primarily relies on the popular Atmospheric scattering model put forward by Koschmieder [7]. The atmospheric scattering model can be mathematically written as $$\begin{equation*}H(k) = A(k)p(k) + \Gamma (1 - p(k))\tag{1}\end{equation*}$$ where H(k) is the raw hazy image, A(k) is haze free image, p(k) is the transmission medium, Γ is the cumulative scattering, and k represents the pixels in the raw hazy image H. Fig. 1 shows the ASM. The main drawback of this method is that it required previous information of the image.