Research On Image Hybrid Segmentation Method Based On Geodesic Active Contour And Fully Convolutional Network

Image segmentation is the process of segmenting an image into disjoint areas with unique properties.It is a key point in the understanding of image scenes and a cornerstone task of computer vision.In recent years,with the continuous deepening of visual scene technology,image segmentation has been widely used in medical diagnosis,autonomous driving,transportation systems,augmented reality and other fields.Traditional segmentation methods based on active contours can evolve a curve to the object contour according to the information of the image itself,and they are often used to quickly segment images with complex structures.However,these methods cannot acquire advanced features by learning a rich supply of data.The fully convolutional network is the deep architecture used to understand key foregrounds.It can learn area feature masks at the pixel level from coarse to fine.Nonetheless,due to the lack of accurate spatial representation and prior information on complex boundaries,it is difficult for the fully convolutional network to improve the ability to recognize contour edge details.Aiming at the difficulties of the above methods in image segmentation,this thesis combines the advantages of geodesic active contour theory and fully convolutional network model,and proposes a hybrid image segmentation method.First of all,this method uses the full convolutional network as the driving framework of the hybrid model.The full convolutional network can extract the high-level features of the image hierarchically and generate feature masks.Then,we construct a differentiable level set layer based on the geodesic active contour theory,and gradually optimize the feature masks on multiple feature channels through the evolution process of the level set.At the same time,the level set reinitialization and Gaussian smoothing operations are integrated into the iterative process of the level set,and packaged as an independent level set layer to achieve a unified training and predicting process.The integrated hybrid segmentation model can optimize the feature mask based on the information of the image itself,and the proposed level set layer can also participate in the data learning process.The experimental research results show that the hybrid segmentation method improves the utilization of low-level features of the image,improves the segmentation details of the target edge,and improves the performance of the segmentation evaluation index.In addition,the proposed level set layer can also be extended to other existing deep segmentation architectures,and can obtain good results.Experimental results on multiple public data sets show that the image hybrid segmentation method can significantly capture the spatial details of the foreground boundary and improve the accuracy of image segmentation.