Deep Learning Model and Dataset for Inspection and Classify Solder Paste Stencil Cleanliness

The inspection of solder paste stencils is a critical step in the production process of electronic devices to prevent issues caused by residual solder paste on the stencils, which can lead to defects in the manufacturing process. This research focuses on selecting an appropriate deep learning model for inspecting and classifying the cleanliness of solder paste stencils using computer vision techniques. We propose the approach for stencil cleanliness inspection using deep learning model along with the dataset. This helps transitioning from manual observation under magnifying devices or handheld digital microscopes (USB) to a system that uses digital images captured by high-resolution cameras. These images are analyzed using deep learning models to classify stencil cleanliness, improving inspection efficiency before the stencils enter the production process. Various deep learning models, including Multilayer Perceptron (MLP), EfficientNetB7, and Vision Transformer, are compared based on performance criteria such as accuracy, recall, and inference time. Additionally, techniques like data augmentation and finetuning are applied to enhance learning efficiency. The dataset includes grayscale images, Sobel-filtered images to enhance edge sharpness, and raw images without preprocessing. The experimental results reveal that the Vision Transformer model achieves exceptional accuracy. The unprocessed raw images yield better classification accuracy than preprocessed images, achieving an accuracy rate of 99.875 %. When compared to the EfficientNetB7 model using Sobel-filtered images for feature extraction [1], which achieved a comparable accuracy of 99.81 %, the Vision Transformer was chosen due to its statistically significantly lower average loss. This indicates that the Vision Transformer is a more suitable model in terms of loss reduction. The findings were supported by a T-test result with a P-value of 0.004.