Research On Endoscopic Image Processing Method Based On Convolution Neural Network

The morbidity and mortality of digestive tract cancer remain high all year round,seriously threatening people’s life and health.Endoscopy,early detection of related diseases,and effective treatment are effective means to reduce the threat of gastrointestinal cancer.Because manual endoscopy relies too much on personal experience,it is prone to problems such as false detection,missed detection,and blurred images.In this paper,the convolutional neural network is used to assist in endoscopic examination to improve the efficiency and accuracy of the examination.Endoscopic images often have defects such as blurring,and the segmentation objects often show the characteristics of small inter-class differences and large intra-class differences,and the size and shape of the lesion area are quite different.Therefore,this paper mainly studies the CNN-based endoscopic Image Deblurring and Semantic Segmentation Methods.In order to solve the difficulties of semantic feature extraction,texture detail reconstruction,and artifacts in endoscopic image deblurring tasks,a UNet deblurring method based on sampling segmentation convolution is designed in this paper.First,in order to fully extract the semantic information of blurred endoscopic images and reconstruct their detailed textures,a novel subsampled segmentation convolution is designed.First,the large-scale features are non-destructively segmented into small-scale feature blocks through equal interval sampling,and then convolution and fusion are performed with the small-scale features.It not only avoids the loss of large-scale feature detail information but also avoids the blurring of small-scale semantic information.Secondly,in order to enhance the detailed reconstruction ability of the network,a feature interaction fusion method is designed.First,use semantic features to activate detailed features,and then fuse the two.Finally,aiming at the difference in brightness and texture of the bright channel,middle channel,and dark channel of the endoscopic image,the gradient reconstruction and frequency domain reconstruction loss functions are designed to improve the sharpness of the reconstructed image.The experimental results on the EAD,Kvasir-SEG datasets,and self-built datasets show that the PSNR of the algorithm in this paper reaches 32.88 d B,33.01 d B,and 32.13 d B,respectively,and the SSIM reaches 0.972,0.973 and 0.9636,respectively,which is better than the mainstream deblurring algorithm.Visually,the texture of the reconstructed image is clearer without artifacts.In order to solve the difficulty of small differences between classes and large differences in the geometry of lesion regions in the segmentation problem,a multi-guided and multi-attention UNet semantic segmentation model is designed in this paper.First,in order to alleviate the information interference between classes,a multi-input encoder is constructed,which inputs the original image multiple times and quickly compresses the space through a large convolution kernel to obtain features with direct spatial dependencies.The features with indirect long-distance dependencies are obtained through deep convolution with small convolution kernels,and the two are compared and fused to achieve complementary enhancement.Second,in order to enhance the size adaptability of the network and improve the segmentation accuracy of complex boundaries,this paper designs multi-stage spatial attention and multi-stage channel attention.Finally,to optimize the feature extraction process,a joint loss function is constructed.The Dice loss and cross-entropy loss of a single decoding stage are combined first,and then the loss functions of multiple stages are combined.On the Kvasir,BKAI-IGH,and self-built esophagus datasets,the m Io U of the algorithm in this paper reached: 0.8301,0.8526,and 0.8493,respectively,which is better than the mainstream segmentation algorithm.This paper takes endoscopic images as the research object,innovates a CNN-based endoscopic image processing method,and designs an endoscopic image deblurring and semantic segmentation network.Compared with the mainstream algorithm,it has achieved better performance,which reflects the theoretical value of the research content of this paper,and is expected to be applied in clinical diagnosis.