Haze severely degrades image quality by suppressing contrast, clarity, and visibility, which is a challenge to tasks in many computer vision applications like autonomous driving, remote sensing, satellite image analysis, video surveillance, and action recognition. These vision-based tasks require clear and highly detailed visual information for efficient analysis and decision-making. This research proposes a novel GAN-based learning method, the Multi-Head Spectral-Attentive Residual Generative Adversarial Network (MHSAR-GAN), to improve the performance of image dehazing. The proposed deep learning-based image haze removal model combines spectral normalization to enhance training stability, multi-head attention for fine-tuning feature selection, and residual learning to retain important structural information to improve single-image dehazing. Depthwise convolutions are also incorporated into the attention mechanism for enhanced spatial feature extraction without added computational complexity. We tested our method on benchmark image dehazing datasets, Haze1K and RESIDE 6K, and compared its performance with state-of-the-art image dehazing models. Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) were pivotal metrics for assessing the model's performance, ensuring a comprehensive evaluation of image quality and structural fidelity. Experimental results indicate that the proposed MHSAR-GAN achieves superior haze removal with preserved fine-grained image details and clearer visibility compared to existing image dehazing methods in quantitative and qualitative comparisons.