Hybrid convolutional vision transformer for extrusion-based 3D food-printing defect classification

Deep learning is generally used to perform remote monitoring of threedimensional (3D) printing results, including extrusion-based 3D food
printing. One of the widely used deep learning algorithms for defect
detection in 3D printing is the convolutional neural network (CNN).
However, the process requires high computational costs and a large dataset.
This research proposes the Con4ViT model, a hybrid model that combines
the strengths of vision transformer with the inherent feature extraction
capabilities of CNN. The locally extracted features in the CNN were merged
using the transformers’ global features with four transformer encoder blocks.
The proposed model has a smaller number of parameters compared to other
lightweight pre-trained deep learning models such as VGG16, VGG19,
EfficientNetB2, InceptionV3, and ResNet50. Thus, the proposed model is
simplified. Simulations were conducted to classify defect and non-defect
images obtained from the printing results of a developed extrusion-based 3D
food printing device. Simulation results showed that the model produced an
accuracy of 95.43%, higher than the state-of-the-art techniques, i.e., VGG16,
VGG19, MobileNetV2, EfficientNetB2, InceptionV3, and ResNet50, with
accuracies of 77.88%, 86.30%, 82.95%, 90.87%, 84.62%, and 93.83%,
respectively. This research shows that the proposed Con4ViT model can be
used for 3D food printing defect detection with high accuracy.