M.A. Thesis:

Digital image dehazing using feature fusion and deep neural network

Fog usually causes severe interference in the image. Such degradation in images poses problems for both To remove the fog effect and extract high-quality images from degraded images, a large number of image de-fogging algorithms with different perspectives have been proposed. Especially in recent years, with the rapid development of deep learning, CNN-based methods have now become the most important image de-fogging methods and have made significant progress in efficient image extraction from foggy images human observers and computer vision systems Considering the very wide application of fog removal in modern systems based on image processing, especially navigation systems and automatic guidance of cars, a high quality fog removal algorithm is proposed in this thesis. The proposed algorithm aims to improve the quality of the output by combining two efficient methods with a mathematical base and adding a deep learning layer In the method proposed in this thesis, first de-fogging is done based on adaptive histogram equalization and gamma correction. This section works in HSV and RGB color space. In the second step, the method based on the color channel is used and by modeling the human brain, which can quickly identify the foggy area from natural landscapes without any additional information from the brightness and saturation of the pixels in a foggy image with Changing the concentration of fog is used to remove fog In the last step, DnCNN deep denoising network was used to achieve a higher quality image. DnCNN network has been used with the justification that after image processing, the residual fog follows a noise pattern and can be considered as noise. Finally, the proposed method and the basic methods were completely shown on the images of two completely standard databases to implement and maintain the quality in the proposed method. In terms of occupied memory, the proposed method is relatively optimal and it can be said that its occupied memory is acceptable.

Keywords: De-fogging, deep learning, histogram equalization, convolutional neural network, color channel