MF-MSCNN: Multi-Feature based Multi-Scale Convolutional Neural Network for Image Dehazing Via Input Transformation

Recently, many deep learning algorithms have been proposed for image dehazing. However, in most of these techniques, the issues of under-exposure and over-saturation are observed. These problems appear because of inadequate consideration of the overall haze level, i.e. Less-Haze (LH), Medium-Haze (MH), and High-Haze (HH), of hazy images while dehazing. Therefore, a Multi-Feature-based Multi-Scale Convolutional Neural Network (MF-MSCNN) is proposed, which considers the haze density of hazy images as a parameter while dehazing. Firstly, the classification operation is performed to categorize the hazy images into LH, MH, and HH. Based on this categorization, a haze density map is generated, which is concatenated to the input hazy image as part of input transformation (IT). Subsequently, the LH, MH, and HH features are extracted using the proposed MF-MSCNN. These features are adaptively chosen by the pooling layer to obtain an efficient transmission map from which the dehazed image is retrieved. The proposed work using IT operation and the MF-MSCNN model produces better results for all categories of hazy images when compared to the existing methods.