Research On Lightweight Dehazing Algorithm And Model Optimization Method Based On Feature Pyramid Network

With the continuous development of science and technology,the explosive growth of data volume and the enhancement of computer computing power,deep learning has made continuous breakthroughs in the field of vision.The traditional deep learning dehazing algorithm only use clear images to guide the training of the dehazing network,but does not use blurred images,thus creating the problem of incomplete dehazing and incomplete detail information.However,the higher the performance of the dehazing model,the larger the number of parameters and the complexity of the model.Therefore,it becomes a key research direction for how to retain more high-frequency and detailed information of the dehazed images,and the lightweight design of the model.For reduce the complexity of the model leading to the degradation of the dehazing performance,this thesis takes the feature pyramid network as the base dehazing structure and investigates the improvement of the dehazing performance and the lightweight design of the model.A dehazing algorithm combining channel attention mechanism,pixel attention mechanism and local residual learning,and using feature adaptive fusion is proposed as the base dehazing model,but the number of model parameters is large.Therefore,the basic dehazing model is improved,and the network model parameters are reduced without affecting the dehazing performance.The model optimization method based on contrastive learning optimizes the basic model and the improved model,and it is verified that the improved dehazing model effectively improves the dehazing effect and reduces the model size.The main research contents of this thesis are as follows:(1)The relationship between the feature adaptive fusion mechanism and the different scales of the pyramid features is investigated,and it is proposed that the feature weights of the different scales of the pyramid assigned by the attention mechanism are fused with the feature information extracted from different layers by the feature adaptive fusion mechanism after the introduction of lateral connections.By adjusting the parameters,the weights of high frequency information and detail information are increased to increase the detail refinement of the dehazed images and improve the dehazing efficiency of the network.(2)The lightweight design of feature pyramid network is studied,the network structure combining attention mechanism and feature pyramid model is improved,the attention mechanism is introduced into the feature fusion stage,the feature adaptive fusion module is replaced,and the number of model parameters is reduced without reducing the dehazing performance.(3)An optimization model of feature pyramid network based on contrast learning is proposed.The training of the dehazing network is guided jointly by contrast learning for both hazy and clear images.The conventional dehazing model uses only clear images and does not utilize hazy images.The contrast learning approach is to bring the image information recovered by the dehazing network closer to the direction of the clear image and farther away from the direction of the hazy image information.This results in enhancing the dehazing performance of the network by making the dehazed image retain more detailed information without increasing the complexity of the model.