Thrips detection (Cap2020)
Study conducted as part of AI4Industry 2026, focusing on the analysis of the impact of controlled noise on improving the detection of insects known as thrips.
Problem Statement
Can image quality be improved, for example, by adding noise to high-quality images?
Methodology
To augment the dataset, controlled noise was applied to high-quality images using several noise models. Gaussian noise was introduced as additive noise following a normal distribution with a mean of 0 and a variance of 0.01. Poisson noise was applied as additive noise correlated with the intensity of each pixel. Salt-and-pepper noise was used as impulsive noise, randomly affecting between 5% and 40% of the image pixels.
Result
At the end of the study, we observed that data augmentation through the controlled introduction of noise into high-quality images has a progressive impact on the F1-score of the thrips detection model.