Simsam: Simple Siamese Representations Based Semantic Affinity Matrix for Unsupervised Image Segmentation

Conventional unsupervised image segmentation methods return many superpixels or object parts and thus tend to over-segmentation. In this paper, we present a novel post-processing approach for unsupervised object-level image segmentation (UnOLIS). Starting with the results of any conventional unsupervised segmentation method, we first combine a global region-based saliency and a robust background feature to cluster the pre-segmented regions into foreground and background. We then design a region growing process, encoded with several object priors, to generate a high quality foreground object segmentation. In parallel, we group the background regions into different stuffs by clustering. We test our method on the Berkeley Segmentation Dataset (BSDS500). Our approach significantly improves conventional unsupervised segmentation methods and achieves almost comparable results as the state-of-the-art supervised image segmentation methods.

Considers the problem of unsupervised multiband image segmentation specifically, using the MDL criterion, an associated complexity measure. According to the MDL principle, the best estimates are those that result in the most compact encoding of the image. An important advantage of such an approach is that it is virtually free of the need to choose arbitrary thresholds, which are typical of many segmentation techniques, and which, in many cases, need to be interactively adjusted in order to get satisfactory results. >

Enormous volumes of image data were generated everyday, but can't be used unless they are organized so as to allow efficient browsing, searching and retrieval. On the other hand, the image data are not only required for developing spatial or geographic data, but they are also critical for military and intelligence applications especially suited to addressing national security, that is, image data are critical for situation awareness and assessment purposes, and they are invaluable for detecting changes and providing relevant information to decision makers. Currently, there is a lack of comprehensive tools that can allow fast and efficient processing of information from huge image data. In order to do automated image retrieval to meet the challenge of organizing and analyzing vast volumes of image data to effectively synthesize the critical information, the key step is image segmentation for the large-scale spatial image. In this paper, one content-based (unsupervised) image segmentation method is proposed and tested out to be very successful for the large-scale image segmentation

A new class of neuro fuzzy systems, based on so-called weighted neural networks (WNN), is introduced and used for unsupervised fuzzy clustering and image segmentation. Incremental and fixed (or grid-partitioned) weighted neural networks are presented and used for this purpose. The WNN algorithm (incremental or grid-partitioned) produces a net, of nodes connected by edges, which reflects and preserves the topology of the input data set. Additional weights, which are proportional to the local densities in the input space, are associated with the resulting nodes and edges to store useful information about the topological relations in the given input data set. A fuzziness factor, proportional to the connectedness of the net, is introduced in the system. A watershed-like procedure is used to cluster the resulting net. The number of resulting clusters is determined by this procedure. Experiments confirm the usefulness and efficiency of the proposed neuro fuzzy systems for image segmentation and, in general, for clustering multi- and high-dimensional data.

This paper proposes an unsupervised color image segmentation approach excellent in multi-texture image segmentation. Actually, it employs a novel texture feature extraction mechanism through the contourlet subband coefficient clustering, which is more effective in image segmentation than the discrete cosine transform based normalization technique (DCT). In addition, it adopts the gradient Bayesian Ying-Yang harmony learning of t-mixtures (BYY-t) for automatic image objects detection so that the image segmentation is in an unsupervised mode. The experiments on the images in Berkeley Segmentation Database and Benchmark (BSDB) database demonstrate the improved performances of this approach in varied and complex color image segmentation. Additional experiments on multi-texture color images further demonstrate its better performances in comparison with those of the state-of-art algorithms.

We present an efficient unsupervised color image segmentation algorithm by combining the local and global color information. By first processing a color image via the proposed color sigma filter, pixels within the same semantic region become more concentrated around their centroid in the perceptual color coordinate system. A k-mean algorithm is then designed to automatically distinguish the image into non-overlapping semantic regions, of which centroids with similar color features are merged automatically. Because of the periodicity in the hue component, we apply two manifolds to completely cover the hue vector, and fuse distinguished regions from both manifolds to obtain the final image segmentation. The computational complexity of our algorithm is O(N), where N is the total number of pixels, and no priori information is assumed. We download sample images from the Internet randomly and apply the proposed algorithm to illustrate the performance of our procedure.

Conditional random field (CRF) has been widely used in optical image and remote sensing image segmentation because of the advantage of directly modeling the posterior distribution and capturing arbitrary dependencies among observations. However, for nonstationary SAR images, applications of CRF often fail because of their nonstationary property. The triplet Markov field (TMF) model is well appropriate for nonstationary SAR image processing, owing to the introduction of an auxiliary field which reflects the nonstationarity. Therefore, we introduce an auxiliary field to describe the nonstationarity of the posterior distribution and propose an unsupervised SAR image segmentation algorithm based on a conditional TMF (CTMF) framework which combines the advantages of both CRF and TMF. The proposed CTMF framework explicitly takes into account the nonstationary property of SAR images, directly models the posterior distribution, and considers the interactions among the observed data. Experimental results on real SAR images validate the effectiveness of the algorithm proposed in this letter.

In this paper, we present an unsupervised color image segmentation algorithm. By first processing a color image via the proposed color sigma filter, pixels within the same semantic region become more concentrated around their centroid in the perceptual color coordinate system. A k-mean algorithm is then designed to automatically differentiate the image into non-overlapping semantic objects. Because of the periodicity in the hue component, we apply two manifolds to completely cover the hue vector, and fuse distinguished regions from both manifolds to obtain the final image segmentation. The computational complexity of our algorithm is O(N), where N is the total number of pixels, and no priori information is assumed. We provide examples to illustrate the performance of our procedure.

A new texture image segmentation algorithm, HMTseg, was recently proposed and applied successfully to supervised segmentation. In this paper, we extend the HMTseg algorithm to unsupervised SAR image segmentation. A multiscale Expectation Maximization (EM) algorithm is used to integrate the parameter estimation and classification into one. Because of the high levels of speckle noise present at fine scales in SAR images, segmentations on coarse scales are more reliable and accurate than those on fine scales. Based on the Hybrid Contextual Labelling Tree (HCLT) model, a weight factor /spl beta/, is introduced to increase the emphasis of context information. Ultimately, a Bayesian interscale and intrascale fusion algorithm is applied to refine raw segmentations.

Low-key images widely exist in imaging-based systems such as space telescopes, medical imaging equipment, machine vision systems. Unsupervised low-key image segmentation is an important process for image analysis or digital measurement in these applications. In this paper, a novel active contour model with the probability density function (PDF) of gamma distribution for image segmentation is proposed. The flexible gamma distribution is used to describe both of the heterogeneous foreground and dark background in a low-key image. Besides, an unsupervised curve initialization method is also designed in this paper, which helps to accelerate the convergence speed of curve evolution. The effectiveness of the proposed algorithm is demonstrated through comparison with the CV model. Finally, an industrial application based on proposed approach is described in this paper.

The traditional multi-resolution Markov random field (MRMRF) model uses two-component Markov random field model on each resolution, and requires training data to estimate the necessary model parameters, which is unsuitable for unsupervised image segmentation. Under this circumstance, a new multi-resolution Markov random field model with variable potential for unsupervised texture image segmentation is presented. The new model solves this problem by introducing a variable potential function for multi-level logistic distribution (MLL) model on each scale. Using this method, the new model can automatically estimate model parameters and produce accurate unsupervised segmentation results. The results obtained on synthetic texture images and remote sensing images demonstrate that a better segmentation is achieved by our model than the traditional MRMRF model.

The application of a Markov random field (MRF) state model in an expectation-maximization (EM)-based approach to unsupervised image segmentation is investigated. In the calculation of the marginal distribution of the state field, it is shown that the use of the expected state values for interacting pixel sites in the computation of the MRF energy function may be interpreted as a mean-field approximation. The implications of calculating a self-consistent expectation of the state field are considered. EM convergence criteria are considered, and a criterion based upon divergence is proposed. Experimental results based on synthetic data illustrate the performance advantage of the mean-field approximation and the computational advantage of using self-consistent expectations. >

In this paper, we study unsupervised Bayesian image segmentation approach which involves capturing model likelihood disparities among different texture features with respect to a global statistical model. Specifically, wavelet-domain hidden Markov models are used to characterize the global textural behavior of images in the wavelet-domain. Three clustering methods, i.e., the K-mean, a soft clustering and a multiscale clustering are studied to convert the unsupervised segmentation problem into the self-supervised process by identifying the reliable training samples. In particular, multiscale clustering involves multiple context models from different scales for context fusion. The simulation results on synthetic mosaics show that the proposed unsupervised segmentation algorithm can achieve high classification accuracy that is close to the supervised one.

While unsupervised segmentation of RGB images has never led to results comparable to supervised segmentation methods, a surprising message of this paper is that unsupervised image segmentation of RGB-D images yields comparable results to supervised segmentation. We propose an unsupervised segmentation algorithm that is carefully crafted to balance the contribution of color and depth features in RGB-D images. The segmentation problem is then formulated as solving the Maximum Weight Independence Set (MWIS) problem. Given superpixels obtained from different layers of a hierarchical segmentation, the saliency of each superpixel is estimated based on balanced combination of features originating from depth, gray level intensity, and texture information. We want to stress four advantages of our method: (1) Its output is a single scale segmentation into meaningful segments of a RGB-D image; (2) The output segmentation contains large as well as small segments correctly representing the objects located in a given scene; (3) Our method does not need any prior knowledge from ground truth images, as is the case for every supervised image segmentation; (4) The computational time is much less than supervised methods. The experimental results show that our unsupervised segmentation method yields comparable results to the recently proposed, supervised segmentation methods [1, 2] on challenging NYU Depth dataset v2.

A novel unsupervised color image segmentation method based on graph cuts with multi-components is proposed, which finds an optimal segmentation of an image by regarding it as an energy minimization problem. First, L*a*b* color space is chosen as color feature, and the multi-scale quaternion Gabor filter is employed to extract texture feature of the given image. Then, the segmentation is formulated in terms of energy minimization with an iterative process based on graph cuts, and the connected regions in each segment are considered as the components of the segment in each iteration. In addition, canny edge detector combined with color gradient is used to remove weak edges in segmentation results with the proposed algorithm. In contrast to previous algorithms, our method could greatly reduce computational complexity during inference procedure by graph cuts. Experimental results demonstrate the promising performance of the proposed method.