1. ABOUT THE DATASET ------------ Title: Trash screen blockage detection using cameras and deep learning: code and dataset Creator(s): Remy Vandaele Organisation(s): Department of Meteorology, University of Reading Rights-holder(s): University of Reading, Crown. Publication Year: 2023 Description: This dataset provides access to the images and network weights produced during our research on trash screen detection, along with minimum working examples allowing to use the network weights on new trash screen camera images. The images come from 54 different cameras with open access feeds provided by the Environment Agency (http://eadevonwebcams.org.uk/), and were collected from January 2022 to January 2023. The images were manually annotated with a "clear" (if the trash screen looks clear), "blocked" (if the trash screen looks blocked) or "other" (if unsure) label. The network weights and minimum working examples allow to estimate labels of new trash screen images using three different methods: a classifier, a siamese network and an anomaly detection method. Cite as: Vandaele, Remy (2023): Trash screen blockage detection using cameras and deep learning: code and dataset. University of Reading. Dataset. https://doi.org/10.17864/1947.000498 Related publication: [1] Vandaele, Remy, Dance, Sarah L and Ojha, Varun (2023). Trash screen blockage detection using cameras and deep learning. ASCE Journal of Computing in Civil Engineering (submitted). Contact: r.a.vandaele@reading.ac.uk, s.l.dance@reading.ac.uk Acknowledgements: The images were provided by the Environment Agency. Prof Sarah Dance and Dr Varun Ojha provided guidance and support on the creation of the dataset and design of the automated labelling methods. 2. TERMS OF USE ------------ Copyright 2023 University of Reading. This dataset is licensed under a Creative Commons Attribution 4.0 International Licence: https://creativecommons.org/licenses/by/4.0/. The images are Crown Copyright licenced under the Open Government Licence v3.0: https://www.nationalarchives.gov.uk/doc/open-government-licence/version/3/. 3. PROJECT AND FUNDING INFORMATION ------------ Title: A FAIR (Flood: Aware, Informed, Resilient) approach to community Flood Risk Dates: August 2022 - January 2023 Funding organisation: Environment Agency Flood and coastal resilience innovation programme (FCRIP) and NERC National Centre for Earth Observation (NCEO) 4. CONTENTS ------------ images/ This repository contains the trash screen images that were used during our research [1]. Each trash screen location is a sub-folder (e.g., images/Cornwall_Crinnis for the Cornwall_Crinnis location). Each location repository is divided into three sub-folders corresponding to the images labels: blocked/, clear/, other/. Note that we provide the full sized images, and that the image crops coordinates used during our research are provided in the file crop_coordinates.txt. weights/ This repository contains the best weights of the networks produced during our research [1]. - classifier.pth contains the weight of the ResNet-50 classification network - siamese.pth contains the weights of the siamese network. - padim_*.pth contains the mean and covariance matrix used for the anomaly detection approach. Minimum working examples on how to use these weights and apply them on new images are respectively provided in classification_network.py, siamese_network.py and padim.py (requires python and pytorch library). 5. METHODS ----------- Please refer to our paper [1] for further details on how this dataset was created.