Ytkarakterisering med maskininlärningsförbättrad bildbehandling

Surface finish is a critical proxy for machining quality, yet conventional contact-based metrology is too slow for in-process control on modern production lines. This thesis proposes an image-driven inspection framework that couples physics-informed pre-processing with deep learning to infer cutting conditions directly from microscope images of milled surfaces. Raw micrographs are first converted to analysis-ready data through adaptive thresholding, morphology-guided noise removal and grayscale-to-height mapping; complementary frequency cues are then extracted via the Fast Fourier Transform (FFT) to capture periodic surface signatures. The resulting feature-rich representations feed a lightweight Convolutional Neural Network (CNN) trained on 4 000 patch-level samples spanning single- and multi-defect scenarios. Compared with stylus profilometry and interferometry ground truth, the network attains 98% classification accuracy on previously unseen data while keeping inference latency compatible with shop-floor cycle times. Patch-based training both mitigates Graphics Processing Unit (GPU) memory limits and implicitly augments the dataset, leading to stable convergence across repeated runs and F1-scores above 0.97 for all defect classes. By delivering reliable, real-time predictions, the system enables closed-loop adjustments of feed rate, tool wear compensation and other cutting parameters, promising reductions in scrap, energy use and re-work for high-value manufacturing sectors such as automotive and aerospace.