Research On Image Dehazing Technology Based On Convolutional Neural Network

This paper studies the image dehazing technology based on convolutional neural network.Image is the basic data of many tasks,but the quality of the image is affected by many factors.Fog in the air is one of them.Foggy weather aggravates the physical phenomenon of light in the process of propagation,resulting in the degradation of image quality.Therefore,the study of image dehazing is very necessary.Traditional image dehazing methods include dehazing on physical models and dehazing on the basis of enhanced images.The effects that can be achieved by traditional methods have certain limitations.Therefore,when deep learning emerged,scholars found that convolutional neural networks can directly extract features from images,and the model complexity is low and the number of weights is small,which can be used in the field of image dehazing.achieve good results.Convolutional neural network is a kind of supervised learning.The learning process uses foggy images to train the network,compares the dehazing results with the non-fog images,and feeds the error back to the network.Fog effect network.Image dehazing technology based on convolutional neural networks is also constantly developing.Some studies choose to deepen neural networks and design more complex convolutional networks to achieve better dehazing effects.Others choose to use other fields.The algorithm is integrated into the convolutional neural network to propose more possibilities.This paper learns from these research results and conducts work in the following three aspects:(1)Design a U-NET dehazing model that combines attention mechanism and wavelet transform.Wavelet transform replaces the pooling layer in the original U-NET,retains more detailed information,and combines the pixel attention mechanism with the channel.The attention mechanism is combined into an attention module,in parallel with the U-NET module,which exists as a feature complement.The model is trained on the OTS dataset of RESIDE,tested on the SOTS dataset,and finally quantitatively compared with other dehazing models.The results show that the model has a better dehazing effect.(2)Design a U-NET dehazing model that combines multi-scale feature fusion module and parallel hole convolution module.The multi-scale feature fusion module refers to the backprojection idea in super-resolution and considers the difference between non-adjacent feature maps.It is hoped to retain more information,and at the same time replace the original "bottleneck" part with the parallel hole convolution module.This module uses 4 hole convolutions with different expansion rates to perform feature extraction in parallel,and in the cascade decoder and The encoder uses the channel attention mechanism.The model is also trained on the OTS dataset of RESIDE,tested on the SOTS dataset,and quantitatively compared and analyzed with other dehazing models.The results verify the effectiveness of the proposed model.(3)A system platform for image dehazing is built and implemented.Based on the two image dehazing models designed in this paper,one can be selected for image dehazing on the system,input a hazy image,and obtain a haze-free image,and It can be compared with the original image to get the performance analysis results.