dc.description.abstract
In this work, a software tool to blur detection was implemented in order to select the best image-frame in a set of key frames. The main objective is to allow the detection and measurement of the blurring level of an image without human intervention, i.e. by artificial intelligence trained to detect blur. During the implementation of this Master Thesis, it was necessary to understand the concept of the blur, the causes and the different algorithms to detect local blur. This work uses multiple methods to detect local blur, analysing neighbour's results with different types of filters. Therefore, the solution is a local blur detector at pixel level that generates two images as output, one mask of blurred/sharped pixel areas, and a grey-image with the different levels of blur per pixel. However, the blurring detection is applied in 1D (one output per single pixel) losing its 2D position in the image but using the neighbouring pixels' information to convert this method in a 1.5D. On the other hand, the decision thresholds to classify the image as blurred or sharp were created by machine learning algorithm based on using Naïve Bayes techniques and Neural Networks solutions, to get a similar result to human blur compression. Finally the result is a stable software able to accomplish the set goals, with an efficiency similar to that of a human person classification.
