Deep Learning Feature And Bipartite Graph Based Image Segmentation

Image segmentation has been developing for many years.For now,image segmentation has a lot of applications in various fields.As a fundamental processing technique in computer vision,image segmentation plays an important role in lots of image processing such as image recognition,target detection.At present,no general survey for the whole scope of image segmentation has been made in the last 10 years,so image segmentation is still a hotspot.The existing image segmentation method is usually based on the manual feature,and the segmentation based on deep learning is more used for image semantic segmentation.This paper is inspired by the deep learning model.Without training the new neural network model,we use existed model to extracts the deep learning features of the image,and cooperate with the multi-scale over-segmentation method to segment the image.In order to ensure the edge smoothness of the final segmentation results and to segment the objects at different scales in the image,the Mean-shift method and the FH image segmentation algorithm are used in this paper.Scale over-segmentation results,where the multi-scale over-partitioned area information is stored by multiple tag matrices.Using the depth learning model to extract the feature map from multiple convolution layers,then according to the regions of over-segmentation,perform a mean pool operation on the feature maps.After that,we combine the Objectness measure and over-segmentation to calculate the complexity of the image,then according to the complexity,determine various distance weight while performing a similarity measure of adjacent regions.At last,we construct a bipartite graph with pixels and over segment regions,and uses the spectral clustering method to achieve segment goal.In order to prove the effectiveness of the proposed algorithm,the BSDS500 dataset and SED dataset are used to test the algorithm and compared the segmentation performance with other algorithms.The experimental results show that the proposed algorithm can segment the complex images more accurately,and show better segmentation performance in PRI,Vo I and BDE 3 indicators.