An Image Raindrop Removal Method Based on Squeeze-Excitation and Residual Fusion

Exceptional efficacy characterizes computer vision in industrial production practices. However, capturing images in adverse weather conditions renders subsequent algorithms susceptible to raindrop interference, leading to failure. This study employs a Generative Adversarial Network (GAN) with attention mechanisms. Initially, morphological opening operation is introduced at the front end of a residual network to heighten mask extraction precision. Furthermore, the Squeeze-and-Excitation block is integrated into the traditional residual network to generate more precise attention maps. Raindrop information derived is effectively eliminated from images through the adversarial training mechanism within the GAN. Experimental evaluations conducted on standard and real-world image datasets demonstrate a significant enhancement in raindrop removal compared to multiple existing algorithms. Results indicate heightened representational capacity within the proposed network model, showcasing superior generalization across diverse scenarios, thereby proficiently achieving raindrop removal while accurately preserving background image details.