Research On Image Segmentation Method Based On Total Variation Spectrum Transform And Deep Convolutional Neural Networ

Image segmentation is one of the core problems in computer vision,which aims to assign a class label to each pixel.It has been widely used in transportation,military,medicine,remote sense and other fields.With the continuous development of hardware devices in various fields,people need to deal with more and more diverse images.However,for small sample image segmentation in ordinary photography or medicine field,traditional methods are often affected by various noises,which results in low segmentation accuracy.For the satellite remote sense image segmentation,the deep learning-based methods can efficiently segment images with large numbers,diverse scales and complex background.Nevertheless,the segmentation accuracy in confusing areas such as target boundary is low due to object shade,insufficient lighting,and small differences within categories.This paper focuses on the applications of total variation spectral transform and deep convolutional neural network in the above two segmentation tasks,and the main work is as follows:Aiming to improve the robustness to noise of traditional image segmentation methods,an image segmentation method based on total variation spectral transform is proposed.Firstly,this method transforms the original image from spatial domain to total variation spectral domain by spectral transform and obtains the multi-scale spectral representation of the original image.Secondly,the separation surface is fitted according to the maximum spectral response time of all pixels,and the band-pass filter is constructed by using the separation surface.Thirdly,the object and background are separated by the inverse transform.Finally,the object structures are refined to obtain final segmentation results.Experiments on the common image and medical image datasets show that the method has high segmentation accuracy in a high noise level and robustness to various noises.To solve the problem that existing semantic segmentation methods have low segmentation accuracy in the confusing regions,the uncertainty and boundary guided network for semantic segmentation is proposed.The network combines edge detection and semantic segmentation to strengthen the feature learning ability.An attention module based on uncertainty is added between encoders and decoders to guide the network focus on the areas with high uncertainty in the encoder feature map.At the same time,the context aggregation module is introduced to realize the boundary guided context space relationship modeling.Experiments on the remote sensing dataset show that this method has satisfactory segmentation performance.