Deep Learning and Machine Learning Approaches for Non Contact Surface Roughness Prediction from Scanning Electron Microscope (SEM) Images
No Thumbnail Available
Date
2025-08-14
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Surface roughness is crucial to the quality, functionality, and durability of manufactured parts.
Nonetheless, there is a problem that conventional methods of contact measurements are costly,
time-consuming, and ineffective in analyzing complicated surface geometry. To address these
constraints, this work proposes deep and traditional machine learning models to specifically
predict surface roughness using Scanning Electron Microscope (SEM) images, eliminating the
need for physical probing or handcrafted surface analysis. Titanium alloys as a target material were
chosen because of the wide range of essential applications, and SEM images were obtained at
various magnification levels to capture the broad range of patterns on their surface. A custom
design Convolutional Neural Network (CNN) was constructed and trained to perform regression
tasks, and two pre-trained deep learning models, ResNet50 and VGG16, were also used.
Simultaneously, the images were used to extract the handcrafted features, which were in turn
trained by the conventional machine learning models, such as Support Vector Regression (SVR),
Random Forest, and Linear Regression. The findings showed that deep learning models mostly
performed better, especially VGG16, than the traditional models by providing higher accuracy and
R2 values above 0.95 in the majority of cases. The developed approach has excellent prospects for
application in the real world to high-precision industries, including aerospace, biomedical, and
electronic manufacturing sectors, allowing for improved performance control and process
optimization speed.
Keywords: Surface Roughness; Surface Roughness Prediction; Scanning Electron Microscope;
Convolutional Neural Network; Support Vector Regression; Random Forest; Linear Regression.