Identifying an Image Classification Model for Welding Defects Detection

This study aims to evaluate three neural network models to identify a base model that will be used for welding defect detection. A methodology is proposed that involves evaluating three different models: EfficientNet, MobilNet, and You Only Look Once, using a dataset of images of dogs and cats for training. The analysis includes training, monitoring, and tuning crucial parameters to achieve an optimal model. The final model selection will be based on accuracy, complexity and training time, ensuring the most accurate and efficient detection of the database (cats and dogs). You Only Look Once is identified as an attractive model due to its high object detection accuracy, crucial for accurate defect identification. Challenges include variability in conditions and defect types, necessitating greater diversity of data sets and model refinement to improve visual presentation. The impact of this study lies in its ability to evaluate the application of convolutional neural networks model for welding defect detection, which can have a great impact on the quality and safety of welded structures in various industrial fields. By achieving more accurate and efficient defect detection, potentially dangerous and costly structural failures can be avoided. Furthermore, the developed methodology and approaches may be applicable to other domains requiring visual anomaly detection, thus extending the impact beyond the realm of welding.