Search

The field of artificial intelligence based image enhancement has been rapidly evolving over the last few years and is able to produce impressive results on non-astronomical images. In this work, we present the first application of Machine Learning based super-resolution (SR) and de-noising (DN) to enhance X-ray images from the European Space Agency’s XMM-Newton telescope. Using XMM-Newton images in band [0.5, 2] keV from the European Photon Imaging Camera pn detector (EPIC-pn), we develop XMM-SuperRes and XMM-DeNoise – deep learning-based models that can generate enhanced SR and DN images from real observations. The models are trained on realistic XMM-Newton simulations such that XMM-SuperRes will output images with two times smaller point-spread function and with improved noise characteristics. The XMM-DeNoise model is trained to produce images with 2.5× the input exposure time from 20 to 50 ks. When tested on real images, DN improves the image quality by 8.2 per cent, as quantified by the global peak-signal-to-noise ratio. These enhanced images allow identification of features that are otherwise hard or impossible to perceive in the original or in filtered/smoothed images with traditional methods. We demonstrate the feasibility of using our deep learning models to enhance XMM-Newton X-ray images to increase their scientific value in a way that could benefit the legacy of the XMM-Newton archive.

Detecting plants in images is central in precision agriculture, but can be challenging due to their small size, similarities in appearance, varying lighting and environmental conditions. Moreover, computational capacity in real-world settings may be limited. This work examines how accurate, computationally efficient real-time plant detection can be achieved on the large and varied benchmarking Open Plant Phenotyping Database, by building upon the State-of-the-Art (SoA) Scaled YOLO v4 real-time object detection model. The effect of pre-processing, namely cropping unnecessary information and increasing contrast, is examined and experimentally shown to improve both accuracy and efficiency. Transfer learning is also leveraged for the deployment of Scaled YOLO v4, using pre-trained weights from the MS COCO data set, and shown to lead to a moderate improvement in accuracy. The proposed final model results in approximately 10% higher accuracy than the existing baseline model, on a representative subset of about half of images in the Open Plant Phenotyping Database. Experiments show that plant detection accuracy is improved for most well represented samples, with errors appearing in particularly challenging cases or caused by data imbalance. This shows the proposed method has significant potential for highly accurate and computationally efficient plant detection in real-world environments.

This work focuses on detecting and localizing a wide range of dynamic textures in video sequences captured by surveillance cameras. Their reliable and robust analysis constitutes a challenging task for traditional computer vision methods, due to barriers like occlusions, the highly non-rigid nature of the moving entities and the complex stochastic nature of their motions. In order to address these issues, a novel hybrid framework is introduced, combining representations on both a local and global scale. A new, handcrafted local binary pattern (LBP)-flow descriptor with Fisher encoding is initially used to effectively capture low level texture dynamics, and a neural network (NN) is deployed after it to obtain a higher level, deeper and more effective representation scheme, capable of robustly discriminating even challenging dynamic texture classes. A novel localization scheme, based on multi-scale superpixel clustering is introduced, in order to detect texture patterns on local and global scales, inside and throughout sequential video frames. Experiments on various challenging benchmark datasets prove our method's efficacy and generality, as remarkable recognition and localization accuracy rates are achieved at a low computational cost, making it appropriate for real world outdoor applications.

Human activity detection from video that is recorded continuously over time has been gaining increasing attention due to its use in applications like security monitoring, smart homes and assisted living setups. The analysis of continuous videos for the detection of specific activities, called Activities of Interest (Aol) in this work, is particularly challenging, as the start and end times of the Aol are unknown, while the Aol themselves feature large anthropometric variations, making their recognition more difficult. Continuously recorded videos also contain periods of inactivity, or activities that are not in the set of Aol, further complicating the problem of detection.This work attempts to overcome these challenges by introducing the concepts of (1) discriminative descriptors, (2) a goal for the Aol, represented by the most discriminative descriptors, (3) a novel, goal based framework for activity detection and recognition in video. We represent AoI goals by descriptors found to be the most discriminative, as defined by a function of their correlation distance from the majority of the data, rather than by semantics or parametric modeling. This ensures flexibility, as activities in video feature many variations which cannot always be adequately represented by model-based approaches. Temporal detection of Aol is based on the distance of the test data from the AoI goals, which is shown to provide accurate results. Activity recognition takes place by applying SVM classification on the detected Aol, and the results are compared with the state-of-the-art on publicly available real world and benchmark research datasets. (C) 2017 Elsevier Inc. All rights reserved.