Study On Smoke Removal Algorithm Of Endoscope Image

In the process of endoscopic surgery,electrocautery or laser ablation of the diseased tissue produces a large amount of chemical smoke,which leads to the obstruction of the surgical field of vision and delays the surgical process,thus increasing the risk of surgery.In recent years,some progress has been made in the research of natural image defogging,but there are still some challenges in the research of medical smoke removal.The causes of smoke in endoscope environment are complex,the scene depth is small,the position of light source is not fixed,and the transmission rate of light source,smoke thickness and other parameters are unknown,so it is impossible to transfer natural image smoke removal methods to medical image smoke removal task effectively.At the same time,due to the lack of paired training datasets in the real environment,the generalization of the model is limited,resulting in artifacts,color distortion and ineffective smoke removal in the real restored images.Therefore,this thesis constructs Smoke Veil Prior Regularized Network（SVPNet）,and verifies the validity and authenticity of the SVPNet model.On this basis,a Multistage Training Dehaze Network（MTDNet）is proposed to solve the problem of using SVPNet model trained on synthetic datasets to test performance degradation on real datasets.After training with this strategy,the model has good generalization,and solves the problem of high concentration smoke occlusion in real environment,providing an effective solution for the study of smoke removal in real medical scenes.The main research content is divided into the following two aspects:（1）Based on the Smoke regularization Prior hypothesis,that is,the endoscope smoke screen is characterized by low channel deviation and low contrast,a smoke removal model SVPNet is constructed using the deep learning method.In this model,smoke-free image,transmission imageand smoke screenare estimated by using a three-layer autoencoder network0).,0).,0)..The total loss function_（10）is designed,which includes image consistency loss function_（89）,self-supervised loss function₀（8） and regularized prior loss function.The experimental results show that the model can accurately estimate the smoke feature map,and,and achieve better smoke removal effect on the synthesized mixed concentration smoke datasets.The background region details of the restored images are clearer and the color layers of the pathological tissue blocks are richer.（2）Due to the difference between synthetic smoke and real smoke,the smoke removal effect of the model trained with the synthetic data is not obvious on the real smoke image.Therefore,a multi-stage training network MTDNet is designed.Stage（a）High concentration synthetic dataset-supervised pre-training;Stage（b）multiple real dataset-unsupervised training optimization,and the supervised loss function₉)and unsupervised loss function₀（67）)corresponding to different stages are designed.The experimental results show MTDNet has obvious smoke removal effect on the synthetic datasets and the high concentration real smoke datasets,and no color distortion and artifacts of the real test image appear.