CNN-Based Classification of Optically Critical Cutting Tools with Complex Geometry: New Insights for CNN-Based Classification Tasks

• Published in: Sensors (Basel, Switzerland) Vol. 25; no. 5; p. 1575
• Main Authors: Bilal, Mühenad, Podishetti, Ranadheer, Girish, Tangirala Sri, Grossmann, Daniel, Bregulla, Markus
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 04.03.2025; MDPI
• Subjects: Accuracy; Artificial intelligence; Automation; Classification; Computer vision; Cutting tools; Datasets; Deep learning; Geometry; Image processing; Light emitting diodes; Lighting; machining tools: Grad-CAM CNN; Mechanization; Methods; neural network performance; Neural networks; ResNet50; sustainability; tool classification; training performance; Germany; machining tools: Grad-CAM CNN; tool classification; ResNet50; training performance; neural network performance; ResNet152; sustainability
• DOI: 10.3390/s25051575
• ISSN: 1424-8220

Sustainability has increasingly emphasized the importance of recycling and repairing materials. Cutting tools, such as milling cutters and drills, play a crucial role due to the high demands placed on products used in CNC machining. As a result, the repair and regrinding of these tools have become more essential. The geometric differences among machining tools determine their specific applications: twist drills have spiral flutes and pointed cutting edges designed for drilling, while end mills feature multiple sharp edges around the shank, making them suitable for milling. Taps and form cutters exhibit unique geometries and cutting-edge shapes, enabling the creation of complex profiles. However, measuring and classifying these tools for repair or regrinding is challenging due to their optical properties and coatings. This research investigates how lighting conditions affect the classification of tools for regrinding, addressing the shortage of skilled workers and the increasing need for automation. This paper compares different training strategies on two unique tool-specific datasets, each containing 36 distinct tools recorded under two lighting conditions—direct diffuse ring lighting and normal daylight. Furthermore, Grad-CAM heatmap analysis provides new insights into relevant classification features.