Research And Implementation Of Complex Maritime Image Enhancement Method Based On Continual Learning

Under complex weather conditions,the light captured by imaging systems is attenuated by small particles in the air,which leads to degradation of the captured images including loss of detail,decreased contrast,and blurry images,affecting visual perception and subsequent computer vision tasks.Therefore,researching image enhancement techniques under complex maritime conditions is of great significance to improve the performance of maritime computer vision tasks.In recent years,with the continuous development of deep learning in visual fields such as object detection,semantic segmentation,and image generation,deep learning has also been applied to image enhancement tasks under complex weather conditions such as rain,haze,and snow.However,such methods often only address image distortion under a single weather condition and cannot deal with the complex weather changes in the maritime environment.In addition,due to the variability of maritime weather conditions,using traditional deep learning methods to achieve complex maritime image enhancement may require frequent adjustment of the neural network model for constantly emerging new weather types,and each adjustment requires integration of all historical training data,which will consume a lot of time and storage resources.To solve the above problems,this thesis conducted the following two research work:(1)Firstly,a general image enhancement module was designed based on the polynomial components of atmospheric scattering model.On this basis,a progressive image dehazing network was proposed,which gradually enhances images by serially connecting several enhancement modules and adding progressive connections between adjacent modules to improve the practicality of the dehazing network.Then,this dehazing network was extended to apply to image deraining and desnowing tasks,achieving a general image enhancement.(2)Then,based on the general image enhancement network,a complex maritime image enhancement method based on continual learning was proposed.When encountering weather types that have not been trained and the model has poor generalization performance,different from the traditional method of expanding training data and retraining the model,this method fine-tunes the model to learn new tasks,while reducing the model’s forgetting of historical tasks through old sample revival method.This makes the image enhancement model capable of enhancing maritime images under various complex weather conditions and quickly compatible with new weather types.Additionally,by introducing image compression and reconstruction networks,cross-attention modules,knowledge distillation loss,and encoding feature classification loss into the continual learning framework,the enhancement effect of complex maritime images was further improved.In this thesis,the proposed methods are compared with several baseline methods on the basis of open-source data and collected data in the field.The experimental results show that the proposed progressive dehazing network outperforms several baseline methods and can be extended to image deraining and desnowing tasks;the proposed complex maritime image enhancement method based on continual learning achieves better continual learning performance than baseline methods.