Research And Application Of Image Segmentation Based On Deep Convolutional Neural Network

As the key technology in the field of image processing and computer vision,image segmentation is widely used in various fields such as industry and medicine.With the rapid development of intelligent industry and smart medical care,the requirements for the accuracy and speed of image segmentation have also increased.The construction of traditional image segmentation models mostly depends on low-level features such as texture and color.When the scene is complex or there are artifacts in the image,the segmentation effect is difficult to improve.In recent years,the Convolutional Neural Network(CNN),which is by virtue of its ability to autonomously extract high-level features of images,has achieved outstanding results in image segmentation tasks of multiple fields.Based on the theory and applications,this dissertation study the segmentation of two-dimensional(2D)infrared thermal images and three-dimensional(3D)medical images.Five improved image segmentation algorithms based on CNN are proposed to establish the multi-dimensional image target region segmentation system.Main tasks as follows:(1)An infrared thermal image target region segmentation algorithm based on 2D fully convolutional network(FCN)and dense conditional random field(DCRF)is proposed.To tackle difficulties of the interested region segmentation in complex background,firstly,the FCN is leveraged for pixel-level features extraction to obtain the coarse segmentation result.Then,the DCRF,which is used to optimize the context information,is perform for detailed segmentation.Five-fold cross-validation experiments were carried out on actually acquired infrared thermal images of the solar panel.Experimental results show that the performance evaluation indicators of the algorithm are all higher than the main existing algorithms.Moreover,this method takes less time and requires less manual interference.In conclusion,the proposed algorithm is capable of the segmentation of the interested region in the infrared thermal image effectively in the complex background.(2)An ischemic stroke segmentation algorithm based on 3D U-Net and dilated convolution is proposed.In order to be able to extract a sufficiently large range of depth-space information from 3D stroke magnetic resonance images(MRI),3D U-Net and the dilated convolution layer are introduced in this dissertation;to solve the common class imbalance problem in medical image segmentation,the experimental data is input to the segmentation network in the form of image sampling blocks.After the network parameters are fixed and the preliminary segmentation results are obtained,the post-processing method is used to refine the segmentation results.This dissertation mainly researched the 3D U-Net algorithm and two algorithms which are based on 3D cascade U-Net and dilated convolution.Experiments were conducted on the 2015 ischemic stroke lesion segmentation(ISLES)challenge dataset.Results show that the introduction of 3D MRI spatial information significantly improves the segmentation performance;the use of image blocks effectively solves the class imbalance problem;the 3D dil.1-Net algorithm based on 3D cascade U-Net and dilated convolution has a test accuracy rate of 0.81.In conclusion,it has high real-time performance and can meet the needs of clinical diagnosis.(3)An ischemic stroke segmentation algorithm based on 3D deep residual U-Net is proposed.To avoid the network degradation problem caused by the deepening of network layers,based on the 3D dil.1-Net algorithm model,this dissertation introduces a residual module in the first-level network of cascaded U-Net.Experimental results show that the accuracy and the hausdorff distance of the 3D deep residual U-Net algorithm on the test set are significantly improved.To some extent,the method can provide an objective basis for clinical diagnosis to a certain extent.To sum up,for the two-dimensional and three-dimensional images,five kinds of target region segmentation algorithms based on deep CNN are proposed,and successfully applied to the infrared natural scene analysis and medical image analysis actually.