Enhancing traffic light recognition in autonomous vehicles: a novel convolutional neural network approach for fog removal using efficient channel attention

Adverse environmental conditions such as haze, rain, and dust can significantly impair visibility for vehicles in motion, leading to blurred vision. Traditional de-fogging techniques often encounter limitations, including reliance on pre-existing assumptions, color distortion, and inadequate real-time performance. Consequently, these methods may fail to effectively distinguish traffic light images post-de-fogging, thereby hindering target recognition in autonomous driving systems. This study introduces a convolutional neural network that integrates deep and shallow feature fusion utilizing the Efficient Channel Attention mechanism. By employing an atmospheric scattering model, shallow features from fogged images are extracted through convolutional layer operations, while deep features are obtained using parallel multi-scale convolution kernels. The integration of shallow and deep features is achieved through a series of fusion steps, enhanced by the application of an effective channel attention mechanism. Ultimately, a transmittance map corresponding to the fog map is derived through nonlinear regression, facilitating the recovery of fog-free images and improving the recognition capabilities of the YOLOv8 model. Experimental results demonstrate that this approach effectively identifies traffic light targets during simulations of vehicles navigating traffic intersections. Furthermore, the quality of fog removal achieved by this method surpasses that of other algorithms, as evidenced by both subjective evaluations and objective analyses.