Deep Learning Dehazing Algorithm Based On Feature Fusion And Image Segmentation

Image acquisition equipment is widely used in production and living fields,which facilitates people’s daily lives and guarantees people’s property safety.Clear images include lots of detailed information,which is one of the important information carriers in today’s society.However,the quality of images captured by the equipment outdoors is easily affected by the weather.The quality of captured images is only good in fine weather,but prone to deterioration in bad weather(such as fog,haze,sand,and dust),resulting in contrast reduction and degradation.Such image degradation problems have different degrees of adverse effects on various computer vision-based processing tasks.To improve the ability of computer vision systems to deal with severe weather,scholars have conducted extensive research on the causes of image degradation in hazy weather and the dehazing algorithm for hazy images.Two dehazing algorithms based on convolutional networks and the atmospheric scattering model are proposed in this paper,the main research contents and achievements are as follows:(1)In this paper,we propose a deep learning dehazing algorithm based on spatial and channel feature fusion,which includes two modules: image dehazing and image enhancement module.The hazy image is first processed by the image dehazing module,which combines the convolution operator and the involution operator for feature extraction.The involution operator is a neural network operator with the opposite properties of the convolution operator.Using the anti-symmetry of the two operators in the space and channel dimensions to fully extract the spatial and channel feature maps of the hazy image and fuse them,then perform sufficient feature extraction on the fused feature maps through multiple feature extraction modules,and then preliminary dehazed image is got by combining the deformed atmospheric scattering model and the hazy image.An image enhancement module is added after the dehazing module,and the CLAHE algorithm is used to enhance the preliminary dehazed images to improve their visual effect and image contrast.(2)Combining the mean square error and the structural similarity index measure,a new loss function is defined to train the neural network,which further reduces the error between the dehazed image of the algorithm and the ground truth,and improves the dehazing effects of the algorithm.(3)In this paper,we propose a deep learning dehazing algorithm based on image segmentation.The core idea of this algorithm is to dehaze the hazy image after image segmentation.In order to reduce the difficulty of estimating model parameters,the input hazy image is segmented before dehazing,and the parameters estimation of the entire hazy image becomes the parameters estimation of each sub-image.The change of haze concentration in each sub-image is smaller compared with the entire image,so its associated parameters are less difficult to estimate.Parameters according to the different hazy conditions in each subimage are obtained,then use those parameters to restore the dehazed image.The final dehazed image will be formed by combining all the sub-dehazed images.In the training process of the convolutional neural network,multiple images are usually input into the network at the same time for batch training,to improve the training efficiency and generalization ability of the network.The dehazing algorithm based on image segmentation presented in this paper evenly divides a hazy image into multiple sub-images for training.For every hazy image,the overall computational load does not increase.Feeding all sub-images of each hazy image into the network simultaneously for training is equivalent to using a larger batch size for parameter training.Therefore,the improved algorithm based on image segmentation can achieve a larger batch size training effect under the same conditions,which improves the training efficiency of the convolutional neural network and the robustness of the algorithm.After training,use a variety of dehazing algorithms and the algorithms proposed by this paper to perform dehazing tasks on several hazy images randomly selected from several data sets.Peak signal-to-noise ratio,structural similarity index measure,and normalized mean square error are used to objectively compare the quality of the dehazed image,and verify the effectiveness and generalization ability of the proposed algorithms.