We propose a novel Mean-Shift method for data clustering, called Robust Mean-Shift (RMS). A new update equation for point iterates is proposed based on the ones of the standard Mean-Shift (MS) and the Blurring Mean-Shift (BMS). Despite its simplicity, the proposed method has not been studied so far. RMS can be set up in both a kernel-based and a nearest-neighbor (NN)-based fashion. Since the update rule of RMS is closer to BMS, the convergence of point iterates is proven based on the BMS convergence theorem of Chen [1]. Experimental results on synthetic and real datasets show that RMS in several cases outperforms MS and BMS in the clustering task. In addition, RMS exhibits larger attraction basins than MS and BMS for identical parametrization; consequently, its kernel variant requires a lower aperture of the kernel function, and its NN variant a lower number of nearest neighbors compared to MS or BMS, to achieve optimal clustering results. In addition, the NN version of RMS does not need to specify a convergence threshold to stop the iterations, contrarily to the NN-BMS algorithm.

We investigated nearest-neighbor density-based clustering for hyperspectral image analysis. Four existing techniques were considered that rely on a K-nearest neighbor (KNN) graph to estimate local density and to propagate labels through algorithm-specific labeling decisions. We first improved two of these techniques, a KNN variant of the density peaks clustering method dpc, and a weighted-mode variant of knnclust, so the four methods use the same input KNN graph and only differ by their labeling rules. We propose two regularization schemes for hyperspectral image analysis: (i) a graph regularization based on mutual nearest neighbors (MNN) prior to clustering to improve cluster discovery in high dimensions; (ii) a spatial regularization to account for correlation between neighboring pixels. We demonstrate the relevance of the proposed methods on synthetic data and hyperspectral images, and show they achieve superior overall performances in most cases, outperforming the state-of-the-art methods by up to 20% in kappa index on real hyperspectral images.

The Hough transform is commonly used for detecting linear features within an image. A line is mapped to a peak within parameter space corresponding to the parameters of the line. By analysing the shape of the peak, or peak locus, within parameter space, it is possible to also use the line Hough transform to detect or analyse arbitrary (non-parametric) curves. It is shown that there is a one-to-one relationship between the curve in image space, and the peak locus in parameter space, enabling the complete curve to be reconstructed from its peak locus. In this paper, we determine the patterns of the peak locus for closed curves (including circles and ellipses), linear segments, inflection points, and corners. It is demonstrated that the curve shape can be simplified by ignoring parts of the peak locus. One such simplification is to derive the convex hull of shapes directly from the representation within the Hough transform. It is also demonstrated that the parameters of elliptical blobs can be measured directly from the Hough transform.

Recent advances in spectral imaging have enabled snapshot acquisition, as a means to mitigate the impracticalities of spectral imaging, e.g., expert operators and cumbersome hardware. Snapshot spectral imaging, e.g., in technologies like spectral filter arrays, has also enabled higher temporal resolution at the expense of the spatio-spectral resolution, allowing for the observation of temporal events. Designing, realising, and deploying such technologies is yet challenging, particularly due to the lack of clear, user-meaningful quality criteria across diverse applications, sensor types, and workflows. Key research gaps include optimising raw image processing from snapshot spectral imagers and assessing spectral image and video quality in ways valuable to end-users, manufacturers, and developers. This paper identifies several challenges and current opportunities. It proposes considering them jointly and suggests creating a new unified snapshot spectral imaging paradigm that would combine new systems and standards, new algorithms, new cost functions, and quality indices.

We propose a new strategy to evaluate the quality of multi and hyperspectral images, from the perspective of human perception. We define the spectral image difference as the overall perceived difference between two spectral images under a set of specified viewing conditions (illuminants). First, we analyze the stability of seven image-difference features across illuminants, by means of an information-theoretic strategy. We demonstrate, in particular, that in the case of common spectral distortions (spectral gamut mapping, spectral compression, spectral reconstruction), chromatic features vary much more than achromatic ones despite considering chromatic adaptation. Then, we propose two computationally efficient spectral image difference metrics and compare them to the results of a subjective visual experiment. A significant improvement is shown over existing metrics such as the widely used root-mean square error.

Hyperspectral (HS) imaging has been used extensively in remote sensing applications like agriculture, forestry, geology and marine science. HS pixel classification is an important task to help identify different classes of materials within a scene, such as different types of crops on a farm. However, this task is significantly hindered by the fact that HS pixels typically form high-dimensional clusters of arbitrary sizes and shapes in the feature space spanned by all spectral channels. This is even more of a challenge when ground truth data is difficult to obtain and when there is no reliable prior information about these clusters (e.g., number, typical shape, intrinsic dimensionality). In this letter, we present a new graph-based clustering approach for hyperspectral data mining that does not require ground truth data nor parameter tuning. It is based on the minimax distance, a measure of similarity between vertices on a graph. Using the silhouette index, we demonstrate that the minimax distance is more suitable to identify clusters in raw hyperspectral data than two other graph-based similarity measures: mutual proximity and shared nearest neighbours. We then introduce the minimax bridgeness-based clustering approach, and we demonstrate that it can discover clusters of interest in hyperspectral data better than comparable approaches.