A deep image segmentation method with deep parameterized priors

Due to the development of convolutional neural networks (CNNs), which have contributed extensively to significant advances in computer vision over the past decade, image segmentation has been profoundly influenced by these powerful network architectures. Currently, the fusion of Convolutional Neural Networks (CNNs) with traditional unsupervised image segmentation techniques remains a popular approach. This paper introduces an innovative framework that achieves unsupervised image segmentation without the need for annotations, leveraging deep clustering techniques. This method proposes a novel depth image segmentation method, which uses depth parameterization prior to image features and segmentation labels. The core idea of our work is to add mask-guided feature generation to reconstruct the original image using a deep neural network. Based on the generated features, a pseudo-label can be generated by converting the feature into a distribution, which then acts as the “base truth” of another deep neural network, splitting the label for each pixel output. In addition, we propose a novel loss function to optimize two networks together and alternately in an iterative process.