Research On Deeplearning For Image Defogging And Under Water Image Enhancement

At present,more and more computer vision systems are widely used in various industries.Most of these computer vision systems work on the premise of inputting clear images.However,in actual scenes,especially in outdoor environments,due to fog,weather and other weather these models ca not be used.The reason is that we cannot guarantee that we can obtain clear and usable image data,therefore image enhancement technology is getting a great deal of attention from researchers.At the same time,with the improvement of our scientific and technological research and development level,the development of the ocean has attracted a plethora attention from all countries.The application of underwater robots for underwater detection is considered an important task,and the application of computer vision is the key to this task.One ring.However,the phenomenon of light absorption and scattering also exists in the underwater environment.Compared with the foggy image,the underwater image will suffer more serious distortion problems,such as reduced contrast and excessive blue-green.Aiming at the impact of foggy and other outdoor environments on image clarity,dehazing image enhancement technology is considered to be an effective method to solve this problem.The dehazing image enhancement technology is usually based on a physical model to estimate the degree of light loss due to absorption and scattering in the atmosphere.This paper proposes a neural network model based on area detection to learn the relationship between the foggy image and the media transmittance in a block-wise manner,and then use the media transmittance map to complete the defogging operation based on the atmospheric scattering model,and enhance the defogging The details of the image.The model is mainly composed of two basic network units and can be trained in an end-to-end manner.One network unit is a network module with a residual structure,which can reduce the optimization difficulty of deep networks;the other network unit is a module with a cascaded cross-channel pooling structure,which integrates multiple semantic levels of fog related features,And improve the ability to express in the nonlinear field.In addition,an evolution-based detail enhancement method has been developed to improve the detail quality of over-smoothing results.The experimental results comparing the synthetic image and the real image show that the method has reached the advanced level in both subjective and objective evaluation standards.In addition,we propose a lightweight version based on the regional neural network model,which can ensure a certain degree of dehazing effect on low-power devices.In order to improve the visual quality of underwater images,we propose an underwater image enhancement model,which can also achieve end-to-end training.The whole model consists of two parts.The first part is a non-linear transformation for preliminary color balance,and the second part is a refinement operation at the channel level to further process the results obtained before.The neural network used is a joint optimization model.The model includes supervised learning and unsupervised learning.Supervised learning is used to align the label value with the predicted value at the pixel level,while unsupervised learning is to further ensure the enhanced Image quality.The experimental results show that the model can well solve the problem of underwater image color cast caused by limited lighting conditions.