Automated blur detection in UAS image datasets utilizing a machine learning approach

Photogrammetry is an increasingly common method of creating high resolution topographic datasets for the purpose of monitoring changes in landscapes over time (among other uses).Image acquisition from Unmanned Aircraft System (UAS) platforms is generally automated, however automatic camera settings don't always result in consistently high image quality (sharpness), potentially necessitating a time-intensive manual review of all images (potentially thousands) before photogrammetric processing.Utilizing machine learning for blur identification in images, our project aims to assist the creation of accurate 3D maps from UAS imaging by automating a quality assessment of large sets of images.By calculating the Laplacian variance of a single channel of an image, we trained a support vector machine (SVM) with more than 2000 images and used our model trained from the SVM to classify images on the level of blurriness.This approach was then integrated with a React based web application, allowing users to submit images, manage projects, and set parameters for the amount of blurriness deemed acceptable for any given project. The blurriness of submitted images is rated (from 0 to 9, low to high) to determine if images are in theacceptable range for the modeling datasets (currently we find that a score of 0-4 is the most reasonable acceptable range) and scored images can be automatically sorted into groups for inclusion in processing runs. Despite early successes with this approach, more work needs to be done to distinguish true blur from the natural optical tendencies of 'problematic' surfaces, e.g. tree canopies and fine-grained sediments.This tool could be used to efficiently sort or classify large image datasets through an open-source web-based approach.