Detecting Sparks from Hot Work Using a Convolutional Neural Network

This thesis investigates the problem of detecting sparks created by hot work using a Convolutional Neural Network (CNN). A dataset is collected for this specific task, using different spark generators, light levels, and distances. The original captured dataset was cropped to create a second dataset with relatively larger sparks. Using these datasets, 6 different models were trained to investigate if pre-training on different datasets would improve accuracy. The results from training are presented, with the best model trained on the cropped dataset correctly identifying sparks an average of 56.4% of the time, and the best model trained on the original dataset correctly identifying sparks an average of 24.1% of the time. The results are discussed, highlighting differences between the datasets and the impact of pre-training. The models are then tested on videos outside of the gathered dataset, and common sources of error are identified and discussed. The discussion is rounded off with a few notes on industrial applications.