Haze removal from a single image is difficult because hazy images contain limited clear information. Traditional methods attempt to estimate haze levels across different image regions using predefined rules and assumptions. In this research, we introduce DehazeNet, a deep-learning model designed to remove haze from images. Built on Convolutional Neural Networks, it leverages the atmospheric scattering model to enhance image clarity. It uses specialized layers that integrate haze-related assumptions and priors. Additionally, it introduces a newly proposed activation function called the Bilateral Rectified Linear Unit, designed to enhance image quality and utilize Maxout units for more effective feature extraction. Another approach explored in this research is GMAN, a deep learning model that differs from DehazeNet. Unlike DehazeNet, GMAN does not rely on estimating haze-related parameters or following an atmospheric model. Instead, it employs an encoder-decoder CNN to reconstruct a clean image directly. GMAN also incorporates residual learning at both fine and coarse scales, enhancing object boundaries and capturing haze patterns more effectively. While traditional methods often darken images or over-sharpen edges, GMAN learns directly from data, avoiding these issues. Quantitative evaluations show that GMAN significantly outperforms DehazeNet (also referred to as HazeNet in our experiments), achieving PSNR scores of 28.59 compared to 22.94, along with superior performance in SSIM and UIQI metrics. These results demonstrate GMAN’s effectiveness in image restoration tasks over traditional and model-based deep learning approaches.