| The presence of rain streaks,raindrops,or water droplets in machine vision images will directly affect the sharpness of the image background,which will greatly affect the performance of the monitoring system.How to effectively remove rain streaks,raindrops or water droplets in an image and improve the quality of the image is an important research topic to be solved.This thesis studies the image de-raining algorithm based on the attentive generative adversarial networks.The specific work includes:1.A general physical model is designed to represent rain streaks and raindrops in the image.This model considers the effect of the refraction and reflection effects of light passing through rain streaks or rain drops on the image.At the same time,in order to obtain a clear background image,the attention map of rain streaks or raindrops is created under the guidance of a binary mask.The information about the background environment is provided to the Generative Adversarial Networks(GAN),which can improve the accuracy of determining whether it is a rainy area.2.A natural image de-raining algorithm based on attentive generative adversarial networks is proposed.The 6-layer attentive recurrent networks with residual connection method and the contextual auto-encoder with symmetric jump connection method improve the efficiency of information flow between network layers and reduce the disappearance of gradient;Visual attention is injected into the generative adversarial network,in order to improve the network’s ability to learn and discriminate the rain streaks or raindrop areas and their surroundings,which can make the background details of the generated rain removal image clearer.Experimental results show that this algorithm is superior to other classical algorithms in subjective and objective aspects of image quality evaluation,and can be effectively applied to the removal of rain streaks and raindrops simultaneously.3.An algorithm for removing water droplets from a rubber conveyor belt image based on an improved attentive generative adversarial networks is proposed.The image morphology operation is added to the attention mechanism,which reduces the gap between the water droplet attention map and the water droplet binary mask map,and improves the sharpness of the generated water droplets removal image.Data normalization is used to avoid the loss of explosion phenomenon,and data augmentation is used to increase the number of training images and improve the generalization ability of the model;By reducing the number of layers of attentive recurrent networks to reduce the number of network sharing parameters,the efficiency of model training and use is improved.The experimental results show that the algorithm can detect the position of water droplets and remove the water droplets,which improves the clarity of the real-time monitoring image of the rubber conveyor belt.And it is superior to the classical algorithms in subjective visual effects and objective quantitative evaluation.The algorithm can be used for safety monitoring of belt conveyors and their belts to ensure the safe operation of equipment and avoid accidents. |
