Combinatorial Energy Learning for Image Segmentation

摘要: We introduce a new machine learning approach for image segmentation that uses neural network to model the conditional energy of given an image. Our approach, combinatorial (CELIS) places particular emphasis on modeling inherent nature dense problems. propose efficient algorithms deep networks function, and local optimization this in space supervoxel agglomerations. extensively evaluate our method publicly available 3-D microscopy dataset with 25 billion voxels ground truth data. On 11 voxel test set, we find improves volumetric reconstruction accuracy by more than 20% as compared two state-of-the-art baseline methods: graph-based output convolutional trained predict boundaries, well random forest classifier agglomerate supervoxels were generated network.