Research On Similar Linear Image Segmentation Method Based On Deep Learning

As one of the basic research directions in the field of computer vision,image segmen-tation has a wide range of applications in automatic driving,intelligent medical treatment,remote sensing road segmentation and other fields.The traditional image segmentation method is to set manual features through the researchers’ prior knowledge,and then di-vide the regions with the same characteristics into the same category.However,the image segmentation method based on deep learning is to extract image features through a neural network,then upsample the deep features to restore the original size,and finally complete the pixel-level classification.Compared with the traditional method,the image segmen-tation method based on deep learning requires a larger amount of data,but has strong anti-noise interference ability and can handle more complex scenes.The current popular deep learning-based image segmentation methods are divided into two structures: one is based on a fully convolutional neural network,and the other is based on an encoder-decoder method.The two methods have different structural advantages,but the steps are sim-ilar.First,image preprocessing,then feature extraction,then up-sampling the extracted features,and finally pixel-level classification.This article summarizes the research on im-age segmentation methods based on deep learning.The main contributions to innovation include the following:Combining the traditional linear filtering method,a fixed convolution kernel module based on linear filtering is proposed to capture the local linearity of retinal blood vessels.At the same time,inspired by the pyramid pooling and strip pooling,a multi-core fusion pooling module based on strip pooling is proposed to capture changes in retinal vessel width and remote context dependence.We added the proposed module to the UNet net-work and constructed a new image segmentation model based on UNet.We apply the pro-posed new UNet-based image segmentation model to three mainstream retinal datasets,and experiments prove that the proposed method is superior to the most advanced meth-ods.Combining the attention mechanism and supervised learning,an attention mecha-nism based on supervised learning is proposed.The channel attention mechanism with position information can improve the performance of the model,but the attention region they focus on is uncertain,and it is difficult to generate an interpretable attention weight map.In this article,we re-calibrate the road image,artificially set the key region of con-cern,perform supervised learning on the attention weight map,and embed the learned position information into the channel attention,so that it deterministic to pay attention to the region we want the network to pay attention to.We combine the proposed super-vised attention mechanism with mainstream network frameworks and apply them to the Deep Globe dataset.Experiments show that the proposed method can greatly improve the segmentation performance of the baseline model.