Search

High-quality image reconstruction techniques allow the generation of high pixel density images from a set of low-resolution micrographs. In general, these techniques consist of two main steps, namely, accurate registration, and formulation of an appropriate forward image model via some restoration method. There exist a wide variety of algorithms to cope with both stages and depending on their practical applications, some methods can outperform others, since they can be sensitive to the assumed data model, noise, drift, etc. When dealing with images generated by Z-contrast scanning transmission electron microscopes, a current trend is based on non-rigid approximations in the registration stage. In our work we aimed at reaching similar accuracy but addressing the most complex calculations in the reconstruction stage, instead of in the registration stage (as the non-rigid approaches do), but using a much smaller number of images. We review some of the most significant methods and address their shortcomings when they are applied to the field of microscopy. Simulated images with known targets will be used to evaluate and compare the main approaches in terms of quality enhancement and computing time. In addition, a procedure to determine the reference image will be proposed to minimise the global drift on the series. The best registration and restoration strategies will be applied to experimental images in order to point up the enhanced capability of this high quality image reconstruction methodology in this field.

Product matching aims at disambiguating descriptions of products belonging to different websites in order to be able to recognize identical elements and to merge the content from those identical items. Most approaches face this matter applying various machine learning methods to textual product descriptions. Recently some authors are including information extracted from an image associated to a textual description of a product. Modern machine learning methods, such as content based information retrieval (CBIR) or deep learning, can be applied to this type of images since they can manage very large data sets for finding hidden structure within them, and for making accurate predictions. This information could boost the performance of the traditional textual matching but at the same time increase the computational complexity of the process. In this paper we review some CBIR and deep learning models and analyse the performance of these approaches when they are applied to images for product matching. The results obtained will help to introduce a combined classifier using textual and image information.

Super-resolution (SR) software-based techniques aim at generating a final image by combining several noisy frames with lower resolution from the same scene. A comparative study on high-resolution high-angle annular dark field images of InAs/GaAs QDs has been carried out in order to evaluate the performance of the SR technique. The obtained SR images present enhanced resolution and higher signal-to-noise (SNR) ratio and sharpness regarding the experimental images. In addition, SR is also applied in the field of strain analysis using digital image processing applications such as geometrical phase analysis and peak pairs analysis. The precision of the strain mappings can be improved when SR methodologies are applied to experimental images.

This paper presents a new approach to model selection based on hypothesis testing. We 4rst describe a procedure to generate di5erent scores for any candidate model from a single sample of training data and then discuss how to apply multiple comparison procedures (MCP) to model selection. MCP statistical tests allow us to compare three or more groups of data while controlling the probability of making at least one Type I error. The complete procedure is illustrated on several model selection tasks, including the determination of the number of hidden units for feed-forward neural networks and the number of kernels for RBF networks. c 2002 Elsevier Science B.V. All rights reserved.

Complexity-penalization strategies are one way to decide on the most appropriate network size in order to address the trade-off between overfitted and underfitted models. In this paper we propose a new penalty term derived from the behaviour of candidate models under noisy conditions that seems to be much more robust against catastrophic overfitting errors that standard techniques. This strategy is applied to several regression problems using polynomial functions, univariate autoregressive models and RBF neural networks. The simulation study at the end of the paper will show that the proposed criterion is extremely competitive when compared to state-of-the-art criteria.

In this paper we describe a pattern recognition system implemented to determine thickness and defocus from HRTEM simulated images. A specific task has been designed to quantify the influence of certain operation parameters of a transmission electron microscope in the global recognition error rate. This influence allows us to estimate human recognition confidence when applying pattern matching to the determination of thickness and defocus from HRTEM maps. The images considered in this task correspond to InP with the sphalerite crystalline structure and were simulated using the EMS computer software package.

The goal of combining the outputs of multiple models is to form an improved meta-model with higher generalization capability than the best single model used in isolation. Most popular ensemble methods do specify neither the number of component models nor their complexity. However, these parameters strongly influence the generalization capability of the meta-model. In this paper we propose an ensemble method which generates a meta-model with optimal values for these parameters. The proposed method suggests using resampling techniques to generate multiple estimations of the generalization error and multiple comparison procedures to select the models that will be combined to form the meta-model. Experimental results show the performance of the model on regression and classification tasks using artificial and real databases.

Estimating Prediction Risk is important for providing a way of computing the expected error for predictions made by a model, but it is also an important tool for model selection. This paper addresses an empirical comparison of model selection techniques based on the Prediction Risk estimation, with particular reference to the structure of nonlinear regularized neural networks. To measure the performance of the different model selection criteria a large-scale small-samples simulation is conducted for feedforward neural networks.

One of the main research concern in neural networks is to find the appropriate network size in order to minimize the trade-off between overfitting and poor approximation. In this paper the choice among different competing models that fit to the same data set is faced when statistical methods for model comparison are applied. The study has been conducted to find a range of models that can work all the same as the cost of complexity varies. If they do not, then the generalization error estimates should be about the same among the set of models. If they do, then the estimates should be different and our job would consist on analyzing pairwise differences between the least generalization error estimate and each one of the range, in order to bound the set of models which might result in an equal performance. This method is illustrated applied to polynomial regression and RBF neural networks.