The paper presents a framework to automatically identify crack patterns and the related features in existing reinforced concrete (RC) bridges. The challenge of this work is to define a tool for detecting the focused defect and highlighting the number and the orientation of cracks, allowing for correct interpretation and driving further evaluations on the residual life of the structure. The study is framed within the increasing interest in monitoring the structural health of existing bridges through automated tools, able to support engineers in the phase of visual assessment and interpretation of structural defects. When dealing with periodic inspection of large bridge portfolios, the support provided by automated tools can be fundamental for planning further strategies aimed at ensuring the structural safety and preventing future disasters. Given a stack of photos of a bridge structural element, an image stitching procedure is proposed to produce a near-complete image of the entire element. On the latter, a pipeline of deep-learning (DL) algorithms is employed to automatically detect and identify cracks (as a combination of object detection and segmentation algorithms). Finally, the proposed tool extracts cracks for counting and defines their orientation (i.e., vertical, horizontal, diagonal), in order to provide near-complete information about the crack pattern for the structural element. A full description of the methodology and the proposed algorithms is reported throughout the manuscript, showing the main pros and cons and assessing the effectiveness of the tool on a real-life case study.

Deep learning (DL) has become the most effective machine learning solution for addressing and accelerating complex problems in various fields, from computer vision and natural language processing to many more. Training well-generalized DL models requires large amounts of data which allows the model to learn the complexity of the task it is being trained to perform. Consequently, performance optimization of the deep-learning models is concentrated on complex architectures with a large number of tunable model parameters, in other words, model-centric techniques. To enable training such large models, significant effort has also gone into high-performance computing and big-data handling. However, adapting DL to tackle specialized domain-related data and problems in real-world settings presents unique challenges that model-centric techniques do not suffice to optimize. In this paper, we tackle the problem of developing DL models for seismic imaging using complex seismic data. We specifically address developing and deploying DL models for salt interpretation using seismic images. Most importantly, we discuss how looking beyond model-centric and leveraging data-centric strategies for optimization of DL model performance was crucial to significantly improve salt interpretation. This technique was also key in developing production quality, robust and generalized models.

Non-alcoholic fatty liver disease (NAFLD) is now the leading cause of chronic liver disease, affecting approximately 30% of people worldwide. Histopathology reading of fibrosis patterns is crucial to diagnosing NAFLD. In particular, separating mild from severe stages corresponds to a critical transition as it correlates with clinical outcomes. Deep Learning for digitized histopathology whole-slide images (WSIs) can reduce high inter- and intra-rater variability. We demonstrate a novel solution to score fibrosis severity on a retrospective cohort of 152 Sirius-Red WSIs, with fibrosis stage annotated at slide level by an expert pathologist. We exploit multiple instance learning and multiple-inferences to address the sparsity of pathological signs. We achieved an accuracy of $ 78.98\pm 5.86\% $, an F1 score of $ 77.99\pm 5.64\%, $ and an AUC of $ 0.87\pm 0.06 $. These results set new state-of-the-art benchmarks for this application.

This paper proposes a novel approach for lossless coding of light field (LF) images based on a macro-pixel (MP) synthesis technique which synthesizes the entire LF image in one step. The reference views used in the synthesis process are selected based on four different view configurations and define the reference LF image. This image is stored as an array of reference MPs which collect one pixel from each reference view, being losslessly encoded as a base layer. A first contribution focuses on a novel network design for view synthesis which synthesizes the entire LF image as an array of synthesized MPs. A second contribution proposes a network model for coding which computes the MP prediction used for lossless encoding of the remaining views as an enhancement layer. Synthesis results show an average distortion of 29.82 dB based on four reference views and up to 36.19 dB based on 25 reference views. Compression results show an average improvement of 29.9% over the traditional lossless image codecs and 9.1% over the state-of-the-art.