A Very Deep Adaptive Convolutional Neural Network (VDACNN) for Image Dehazing

The main challenge faced by the existing methods is that they cannot efficiently eliminate the haze from the dense hazy or foggy images. The haze features of dense hazy images are not effectively learnt by the Deep Neural Networks. To resolve this drawback, a very deep adaptive convolutional neural network model is proposed for efficient image dehazing. The hazy images are first categorized into two categories viz., Less-Haze and High-Haze. Two Very Deep Convolutional Neural Networks (VDCNNs), viz., Less Haze-VDCNN and High Haze-VDCNN are developed separately for the classified images. Then the Less Haze-VDCNN is trained using the input hazy images that are less haze and their transmission maps as output. Similarly, the High Haze-VDCNN is trained separately with high-hazy images and their transmission maps. After the training process, a hazy image can be adaptively dehazed from one of the two trained VDCNNs based on the hazy image categorization. The proposed VDACNN exhibits better results for dense hazy images in comparison to existing approaches.