Research On Underwater Optical Image Enhancement Method Based On Image Fusion And Deep Network Model

Underwater optical imaging technology is one of the main means and tools of ocean exploration,which can capture underwater information more accurately.However,due to the complex underwater environments,underwater optical imaging faces many challenges.One of the biggest challenges is the light scattering and absorption effects in the underwater environments.These negative factors lead to widespread contrast reduction,color distortion and blurred details in underwater images,which affects the visual quality of images.Therefore,how to effectively improve the quality and clarity of underwater optical images has become one of the problems concerned by academia and industry.This thesis deeply analyzes a series of problems in the field of underwater optical image enhancement,including color scattering caused by wavelength dependent scattering,limited nonlinear capability of CNN model in dynamic turbidity scenes,unstable prior knowledge in non-uniform illumination scenes,and pathological modeling of degraded processes in different water types.An underwater optical image enhancement method based on image fusion and deep network model is studied.(1)Aiming at the problem of color scattering caused by wavelength dependent scattering,an adaptive weighted underwater optical image enhancement method based on information balance is proposed.Firstly,a white balance method based on the main color channel is designed to compensate the attenuation of light energy to improve the color distortion.Secondly,an adaptive image fusion method based on information balance is proposed.By quantifying the relationship between the weight map and the quality of fused image,the estimation of weight map is modeled as an optimization problem.According to the quality evaluation results of the fused image,the objective function is constructed to control the optimization process of weight map,and the particle swarm optimization algorithm is used for iterative solution.Experimental results show that the proposed method can effectively solve the color distortion and contrast reduction caused by color scattering.(2)Aiming at the strong scattering problem in dynamic turbidity scene,an underwater optical image enhancement method based on multi-branch convolutional neural network is proposed.By studying the dynamic convolution mechanism,an attention-guided dynamic multi-branch block is proposed,which consists of a dynamic feature selection module and a multi-scale channel attention module.The dynamic feature selection module extracts multiscale features by using different receptive fields,and designs a nonlinear strategy to realize dynamic fusion of multi-scale features by introducing attention mechanism.The multi-scale channel attention module uses the channel attention mechanism to learn the correlation of channel dimensions in different branches,and then extracts discriminant features.Experimental results show that the proposed method can effectively deal with dynamic turbid scenes and obtain high quality output images.(3)Aiming at the limitation problem of prior knowledge instability in non-uniform illumination scene,an underwater optical image enhancement method based on color balance and luminance compensation is proposed.Firstly,an adaptive gamma correction technique is constructed to eliminate the non-uniform color fading problem.Then,in order to solve the problem of blur and noise,the image decomposition algorithm is introduced.The color corrected image is regarded as the linear superposition of texture layer and structure layer,and then the information of each layer of image is processed,respectively.In the structure layer,a local luminance adaptive compensation function based on the normalized model is constructed to improve the luminance distortion problem.The smooth weight coefficient is introduced in the texture layer to enhance the detail information and realize the purpose of removing noise.Finally,the enhanced structure layer and texture layer is fused together to obtain the final output.Experimental results show that the proposed method can effectively suppress the interference of non-uniform illumination and improve the visibility of underwater images.(4)Aiming at the pathological modeling problem of image degradation process caused by different attenuation parameters in different water types,an underwater optical image enhancement with degradation-aware vision transformer is proposed.This method deeply analyzes the imaging characteristics of underwater images,which combines the light absorption and scattering characteristics with the deep network model,and realizes the end-to-end underwater image enhancement through the encoder-decoder structure.Firstly,a space-channel Transformer is constructed so that the model can assign greater weight to space areas and color channels with serious quality degradation.Secondly,a feature interaction module is proposed,which can extract more discriminant expression features by introducing transmission map.In addition,the progressive multi-stage learning strategy is introduced to constraint the convergence process of model.Experimental results show that the proposed method can significantly correct color bias and improve image contrast and texture information when applied to multiple attenuated underwater scenes.In summary,aiming at the main problems existing in current underwater optical image enhancement methods,this thesis designs four underwater optical image enhancement methods using image fusion and depth network model to effectively improve the visualization effect and clarity of underwater images.The research work in this thesis can not only enrich the existing relevant technologies,but more importantly,it is conducive to the application of computer vision research results to the practical engineering of underwater robot autonomy and precision fishing.Therefore,the research content of this thesis has important theoretical significance and practical application value.