The ubiquitous presence of Internet-of-Things devices represents a menace to cybersecurity since their low computational power does not allow classic countermeasures techniques. In recent years, anomaly detection using Machine Learning has acquired popularity among researchers; however, current datasets in literature are sub-optimal, making exhaustive benchmarks challenging to create. Moreover, the current proposed ML solutions do not consider the necessity of adapting to IoT networks that mutate fast. We propose a Continuous Learning approach to anomaly detection on IoT networks: frequent model retraining and the exploitation of previous knowledge allow a classifier to quickly adapt to new network configurations. We deployed this approach firstly by defining a benchmark methodology for ML algorithms that helps train and test models for anomaly detection in IoT networks. This phase allowed us to identify the best-performing algorithm for the task. Then, we employed Semi-supervised Learning techniques to deploy the continuous part concretely. Our results show that applying Continuous Learning can progressively improve anomaly detection performance in key metrics such as Recall (up to 75%) and Roc AUC (up to 33%).

Recently, the so-called cell-free (CF) massive multiple-input multiple-output (MIMO) architecture has been introduced, wherein a very large number of distributed access points (APs) simultaneously and jointly serve a much smaller number of mobile stations (MSs). The paper extends the CF approach to the case in which both the APs and the MSs are equipped with multiple antennas, proposing a beamfoming scheme that, relying on the zero-forcing strategy, does not require channel estimation at the MSs. We contrast the originally proposed formulation of CF massive MIMO with a user-centric (UC) approach wherein each MS is served only by a limited number of APs. Exploiting the framework of successive lower-bound maximization, the paper also proposes and analyzes power allocation strategies aimed at either sum-rate maximization or minimum-rate maximization, both for the uplink and downlink. Results show that the UC approach, which requires smaller backhaul overhead and is more scalable that the CF deployment, also achieves generally better performance than the CF approach for the vast majority of the users, especially on the uplink.

This paper describes the most recent achievements in speckle reduction of COSMO-SkyMed (CSK@) synthetic aperture radar (SAR) data. An advanced multiresolution despeckling filter, based on undecimated wavelet transform (UDWT) and maximum a-posteriori (MAP) estimation has been specialized and optimized to CSKê data, both single- and multi-look. The tradeoff between performances and computational complexity has been investigated: Laplacian-Gaussian and generalized Gaussian (GG) priors for MAP estimation in UDWT domain differ by one order of magnitude in computation cost. Pre-processing of point targets and segmentation of wavelet planes has been exploited to effectively handle the heterogeneity of the data. Besides traditional supervised methods to evaluate the quality of despeckling, a novel procedure, fully automated, based on bivariate analysis of noisy and denoised image has been devised.

In this work, maximum a posteriori (MAP) despeckling, implemented in the multiresolution domain defined by the undecimated discrete wavelet transform (UDWT), will carried out on very high resolution (VHR) SAR images and compared with earlier multiresolution approaches developed by the authors. The MAP solution in UDWT domain has been specialized to SAR imagery. Every UDWT subband is segmented into statistically homogeneous segments and one generalized Gaussian (GG) PDF (variance and shape factor) is estimated for each segment. This solution allows to effectively handle scene heterogeneity as imaged by the VHR SAR system. Segmentation exploits a Tree Structured Markov Random Field (TSMRF), which is a low complexity MRF segmentation that allows the estimation of the number of segments and the segmentation itself to be carried out at same time. Experiments performed on a single-look VHR X-band SAR images demonstrate that the segmented approach is effective whenever the classical circular Gaussian model of complex reflectivity may no longer hold.

In the field of handwriting recognition, classifier combination received much more interest than the study of powerful individual classifiers. This is mainly due to the enormous variability among the patterns to be classified, that typically requires the definition of complex high dimensional feature spaces: as the overall complexity increases, the risk of inconsistency in the decision of the classifier increases as well. In this framework, we propose a new combining method based on the use of a Bayesian Network. In particular, we suggest to reformulate the classifier combination problem as a pattern recognition one in which each input pattern is associated to a feature vector composed by the output of the classifiers to be combined. A Bayesian Network is then used to automatically infer the probability distribution for each class and eventually to perform the final classification. Experiments have been performed by using two different pools of classifiers, namely an ensemble of Learning Vector Quantization neural networks and an ensemble of Back Propagation neural networks, and handwritten specimen from the UCI Machine Learning Repository. The obtained performance has been compared with those exhibited by multi-classifier systems adopting the classifiers, but three of the most effective and widely used combining rules: the Majority Vote, the Weighted Majority Vote and the Borda Count.

Mammography is a not invasive diagnostic technique widely used for early cancer detection in women breast. A particularly significant clue of such disease is the presence of clusters of microcalcifications. The automatic detection and classification of such clusters is a very difficult task because of the small size of the microcalcifications and of the poor quality of the digital mammograms. In literature, all the proposed methods for the automatic detection focus on the single microcalcification. In this paper, an approach that moves the final decision on the regions identified by the segmentation in the phase of clustering is proposed. To this aim, the output of a classifier on the single microcalcifications is used as input data in a clustering algorithms which produce the detected clusters. As final output the system highlights the suspicious clusters, leaving to the specialist the diagnosis responsibility. The approach has been successfully tested on a standard database of 40 mammographic images, publicly available.

In the framework of a saliency-based approach for segmenting cursive handwriting into elementary strokes, we propose a smoothing technique based on the use of a Wavelet Transform to describe the electronic ink at different resolutions. According to such an approach, derived in analogy with those proposed in the literature for early visual tasks in primates, curvature maxima corresponding to actual segmentation points are separated from those produced by the different source of noise affecting the handwriting generation process. This is obtained by evaluating the curvature maxima at different levels of resolution, and by arranging them into a pyramidal structure, from which a saliency map is eventually achieved by combining the values across all possible scales. Such maps enjoy the property of exhibiting higher values in correspondence of regions of the original ink where curvature maxima survive along difference scales, thus indicating that those regions should correspond to the parts of the ink where two strokes join. Experiments performed by using both standard databases and one of online words collected in our laboratory, have shown high performance in terms of correct segmentation, stability and predictability of the results.

Segmentation algorithms are often used in many image processing applications like compression, restoration, content extraction, and classification. In particular as for content extraction works carried out in the past decade have demonstrated that multi-frequency fully polarimetric SAR observations are particularly interesting, thanks to physical properties of the backscattered signal at various frequencies and polarizations. To achieve a good classification, the main difficulty is that SAR images are often embedded in heavy speckle. Segmentation of multi/hyperspectral (optical) imagery is obtained by means of algorithms based on image models, which exploit the spatial dependencies of land-covers. Unfortunately, speckle noise hides such spatial dependencies in observed SAR data. With the aim of investigating on a content extraction algorithm capable of discriminating cover classes present in the observed SAR image, heterogeneity features are used here to emphasize spatial dependencies in the data. Thus, observed pixel values are mapped into features, that take "similar" values on "similar" textures. This allows for using the same procedure of the optical case. Obviously, homogeneity/heterogeneity feature and segmentation quality are fundamental for classification accuracy. Here, the problem is tackled through the joint use of information theoretic SAR features and of a segmentation algorithm based on Markov Random Fields (MRFs).

Segmentation algorithms are often used in many image processing applications like compression, restoration, content extraction, and classification. In particular as for the content extraction, works carried out in the past decade have demonstrated that multi-frequency fully polarimetric SAR observations content are particularly interesting, thanks to physical properties of the backscattered signal at various frequencies and polarizations. To achieve a good classification, the main difficulty is that SAR images are often embedded in heavy speckle. Segmentation of multi/hyperspectral (optical) imagery is obtained by means of algorithms based on image models, which exploit the spatial dependencies of landcovers. Unfortunately, speckle noise hides such spatial dependencies in observed SAR data. With the aim of investigating on a content extraction algorithm capable of discriminating cover classes present in the observed SAR image, homogeneity/heterogeneity features are used here to emphasize spatial dependencies in the data. Thus, observed pixel values are mapped into features, that take "similar" values on "similar" textures. This allows for using the same procedure of the optical case. Obviously, homogeneity/heterogeneity feature and segmentation quality are fundamental for classification accuracy. Here, the problem is tackled through the joint use of information-theoretic SAR features and of a segmentation algorithm based on Markov Random Fields (MRFs).

In this paper the application of a transform coding technique, based on overcomplete independent component analysis (ICA), for the compression of single look intensity synthetic aperture radar (SAR) images is explored. The method has the advantage of representing the image through almost statistically independent coefficients, with an assigned distribution, so that a scalar entropy constrained quantizer, optimized for the coefficients statistics, can be used. Numerical results on ERS-1 data are presented.

The forest cover classification is extremely important for land use planning and management. In this framework, the application of pixel based classifications of middle resolution images is well assessed while the usefulness of segmentation processes and object classification is still improving. In this paper, a method based on tree-structured Markov random field (TS-MRF) is applied to Landsat TM images in order to assess the capability of the TS-MRF segmentation algorithm for discriminating forest-non forest covers in a test area located in the Eastern Italian Alps of Trentino. In particular, the regions of interest are selected from the image using a two step process based on a segmentation algorithm and an analysis process. The segmentation is achieved applying a MRF a-prior model, which takes into account the spatial dependencies in the image, and the TS-MRF optimisation algorithm which segments recursively the image in smaller regions using a binary tree structure. The analysis process links to each object identified by the segmentation a set of features related to the geometry (like shape, smoothness, etc.), to the spectral signature and to the neighbour regions (contextual features). These features were used in this study for classifying each object as forest or non-forest thought a simple supervised classification algorithm based on a thresholds built on the feature values obtained from a set of training objects. This method already allowed the detection of the forest area within the study area with an accuracy of 90%, while better performances could be achieved using more sophisticated classification algorithm, like Neural Networks and Support Vector Machine.

SAR images are affected by speckle that affects radiometric resolution and class discrimination capabilities. Recently, different speckle reduction techniques based on maximum a posteriori (MAP) estimation have been proven to have very good performances. These techniques are based on the introduction of an a priori statistical model of the speckle free image to be estimated. We propose a MAP method using more than one sub-band filtered intensity images and a Markov random field (MRF) a priori model. The method has been experimented on simulated and real images

We present a fast Bayesian algorithm for the segmentation of remote-sensing images. It alternates two processing steps, the binary Bayesian segmentation of regions, and the separation of non-connected same-class regions, which both present relatively low complexity. As a result, a detailed and reliable K-region segmentation map can be obtained in limited CPU-time. In addition, the map is organized in a tree-structure (not necessarily binary) which helps gaining insight about the meaning of component regions.

The Bayesian/MRF approach guarantees high-quality image segmentation, at the price of a significant computational cost. To limit complexity, one can resort to the recently proposed tree-structured MRF-based segmentation, which converts a Kclass problem in a sequence of much simpler binary tasks. The binary-tree structure imposed to the image, however, can also reduce segmentation accuracy. Here we propose an improved tree-structured segmentation algorithm, where disjoint regions, even belonging to the same class, are immediately split, giving rise to a generic tree structure, and are grouped again in classes only at the end of the algorithm. As a result, both segmentation speed and accuracy increase.

We present a new image segmentation algorithm based on a tree-structured binary MRF model. The image is recursively segmented in smaller and smaller regions until a stopping condition, local to each region, is met. Each elementary binary segmentation is obtained as the solution of a MAP estimation problem, with the region prior modeled as an MRF. Since only binary fields are used, and thanks to the tree structure, the algorithm is quite fast, and allows one to address the cluster validation problem in a seamless way. In addition, all field parameters are estimated locally, allowing for some spatial adaptivity. To improve segmentation accuracy, a split-and-merge procedure is also developed and a spatially adaptive MRF model is used. Numerical experiments on multispectral images show that the proposed algorithm is much faster than a similar reference algorithm based on "flat" MRF models, and its performance, in terms of segmentation accuracy and map smoothness, is comparable or even superior.

Because of its computational complexity, vector quantization (VQ) can work only on relatively small vectors, which severely limits its encoding performance. Trellis-coded VQ (TCVQ) circumvents this problem by using VQ in combination with a trellis encoding strategy, thus treating very large blocks of data at once. TCVQ is based on a size-N VQ codebook which is recursively partitioned to form two (possibly nearly optimal) codebooks of size N/2, four of size N/4, etc. Here we show that such a tree of codebooks can be easily designed, without the need of any postprocessing, by means of the Kohonen algorithm. Numerical experiments show the effectiveness of the proposed approach.

There is an ever-growing interest in the compression of SAR data because of the huge resources required for storage and transmission. This is especially true for spaceborne sensors, given the limited capacity of the downlink channel. Unfortunately, SAR data lack the useful properties on which compression algorithms rely; indeed, these are present in the focused images, but focusing is too complex for on-board implementation at this time. Poggi et al. (2000) proposed to perform on the satellite only the low-complexity range focusing, which increases the data correlation and better concentrates their energy. These properties were then exploited by adopting a variable-rate vector quantizer, with a clear performance improvement with respect to reference techniques. However, vector quantization (VQ) is too complex for actual on-board implementation, and therefore, here we replace VQ with trellis-coded VQ. To limit complexity, only small vectors are used, which reduces VQ's ability to exploit data dependencies; on the other hand, trellis coding allows one to encode large blocks of data at once, and to obtain a better partition of the input space. Experiments on real SAR data show that the overall performance is comparable to that of Poggi et al., but the complexity is much lower, making on-board implementation possible.