Similarity analysis on 3D CAD models with convolutional neural networks

To perform similarity analysis on 3D CAD (Computer-Aided Design) models, this study presents a detailed examination and comparative evaluation of several deep learning-based Convolutional Neural Network (CNN) architectures, namely ResNet50, EfficientNet, Xception, MobileNetV2, InceptionResNetV2, and DenseNet121. While traditional methods often suffer from limited accuracy and high computational cost, CNN architectures offer advanced feature extraction capabilities, enabling more efficient and effective similarity analysis in industrial design and manufacturing workflows. The dataset constructed for this research consists of images generated from each 3D CAD model at seven distinct viewing angles 0°, 45°, 90°, 135°, 180°, 225°, and 270° along the x, y, and z axes. This multi-view imaging approach allowed for a comprehensive evaluation of each architecture's performance from different perspectives, providing a robust basis for assessing their overall effectiveness in similarity detection tasks. The variation in performance across different viewing angles for each CNN architecture highlighted their respective strengths and weaknesses in visual similarity analysis. Angle-dependent accuracy fluctuations were treated as a key evaluation criterion, enabling a multi-faceted assessment of the models. Additionally, an analysis of model complexity and parameter efficiency revealed that higher parameter counts do not necessarily lead to superior performance. Efficient and optimized architectures were shown to achieve competitive results with significantly lower resource requirements. This study demonstrates the feasibility of integrating CNN-based approaches into similarity analysis processes within industrial design. The proposed visual similarity framework facilitates rapid and meaningful access to prior designs during the early stages of development, thereby reducing project timelines, minimizing redundant research, and optimizing resource utilization. Ultimately, deep learning-driven analysis is positioned as a strategic enabler of innovation in industrial design and manufacturing processes.