Estimating Infrared-to-Visible Light Ratio in Images Using Gradient-Boosted Decision Trees and Multilayer Perceptrons
(2025) In Master's Theses in Mathematical Sciences FMAM05 20242Mathematics (Faculty of Engineering)
- Abstract
- This thesis investigates using machine learning for estimating the infrared-to-visible light ratio (qIR) in images, which is used for day/night synchronization in surveillance cameras. Three different models were developed, namely a regression multilayer perceptron (RMLP), a regression-by-classification MLP (RbC MLP), as well as a gradient-boosted decision tree (GBDT). Using an internal dataset collected in both laboratory and real environments, the MLP models had the best performance, with both the RMLP (RMSE = 0.071, R2 = 0.883) and the RbC MLP (81.2% accuracy) showing good results. The GBDT model showed limited capabilities of generalizing to unseen data (RMSE = 0.137, R2 = 0.592). An ablation study underscored the importance of using... (More)
- This thesis investigates using machine learning for estimating the infrared-to-visible light ratio (qIR) in images, which is used for day/night synchronization in surveillance cameras. Three different models were developed, namely a regression multilayer perceptron (RMLP), a regression-by-classification MLP (RbC MLP), as well as a gradient-boosted decision tree (GBDT). Using an internal dataset collected in both laboratory and real environments, the MLP models had the best performance, with both the RMLP (RMSE = 0.071, R2 = 0.883) and the RbC MLP (81.2% accuracy) showing good results. The GBDT model showed limited capabilities of generalizing to unseen data (RMSE = 0.137, R2 = 0.592). An ablation study underscored the importance of using both data augmentation and regularization, which improved the overall accuracy of the RbC MLP by approximately 48 percentage points. Overall, the results demonstrate the potential of using multilayer perceptrons for spectral estimation in surveillance cameras. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9185839
- author
- Ahlin, Axel LU
- supervisor
-
- Viktor Larsson LU
- Ludvig Dillén LU
- organization
- course
- FMAM05 20242
- year
- 2025
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- machine learning, infrared light, surveillance cameras, multilayer perceptron, gradient boosting, regression-by-classification, color histograms, spectral estimation
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUTFMA-3569-2025
- ISSN
- 1404-6342
- other publication id
- 2025:E10
- language
- English
- id
- 9185839
- date added to LUP
- 2025-03-04 15:20:01
- date last changed
- 2025-03-04 15:20:01
@misc{9185839, abstract = {{This thesis investigates using machine learning for estimating the infrared-to-visible light ratio (qIR) in images, which is used for day/night synchronization in surveillance cameras. Three different models were developed, namely a regression multilayer perceptron (RMLP), a regression-by-classification MLP (RbC MLP), as well as a gradient-boosted decision tree (GBDT). Using an internal dataset collected in both laboratory and real environments, the MLP models had the best performance, with both the RMLP (RMSE = 0.071, R2 = 0.883) and the RbC MLP (81.2% accuracy) showing good results. The GBDT model showed limited capabilities of generalizing to unseen data (RMSE = 0.137, R2 = 0.592). An ablation study underscored the importance of using both data augmentation and regularization, which improved the overall accuracy of the RbC MLP by approximately 48 percentage points. Overall, the results demonstrate the potential of using multilayer perceptrons for spectral estimation in surveillance cameras.}}, author = {{Ahlin, Axel}}, issn = {{1404-6342}}, language = {{eng}}, note = {{Student Paper}}, series = {{Master's Theses in Mathematical Sciences}}, title = {{Estimating Infrared-to-Visible Light Ratio in Images Using Gradient-Boosted Decision Trees and Multilayer Perceptrons}}, year = {{2025}}, }