An efficient deep learning based fog removal model for multimedia applications


Abstract: In the present era of technology, several applications such as surveillances systems, security and object recognitions mainly depend on the contents of an image. In this context, the hazy/foggy environment and/or other adverse climatic conditions degrade the image contents that severely influences the result of related applications. The effective haze removal from a single image decides the reliability of these systems. The convolutional neural network (CNN) based techniques are widely used among the available image dehazing methods. However, in CNN based image dehazing techniques, the robustness and accuracy of the learning models are based on the improvement of transmission estimation without giving much concern to the atmospheric light. Therefore, in this paper, the accurate and efficient deep CNN based image dehazing model, which take care the minute information elements during the learning of feature map, is proposed. Besides, the proposed model handles the hallo, blocking artifacts, retainment of fine edges, white region handling, and color fidelity problems, which are primarily responsible for image sharpening and structural stability. For the evaluation of proposed method, the extensive experiments on synthetic and real world images are performed using existing and proposed techniques. The qualitative and quantitative analysis of experimental result shows that the proposed model is more efficient over the existing prior-based and learning-based methods.

Keywords: Single image dehazing, atmospheric light, convolution neural network, transmissivity, image restoration

Full Text: PDF