Defect detection dataset

The software development industry has experienced a gradual yet apparent transformation. Developers are attempting to respond to this growing requirement by adding automation to all that is becoming more and more a matter of software. Techniques such as pipelining as continuous integration (CI) and continuous delivery (CD) have evolved much due to the massive demands of the release and portability of new features and applications. Concurrently, quality control of industry, especially surface inspection, encounters difficulties with manual inspection or traditional machine learning, which is slow, prone to error and hard to scale. The application of AI models into production presents further complexity since there is no consistency in the environment, problems with reproducibility, and integration challenges. Objective: The proposed study is expected to create a scalable, reproduceable, and low-code AI implementation framework that consists of CI\/CD pipelines to handle automated model management, real-time inference, and adaptive retraining. Methodology: The framework was tested with three datasets: NEU Surface Defect Database (five types of defects: crazing, inclusion, scratches, pitted surface and rolled-in scale), CIFAR-10 dataset (general-purpose image classification) and Intel Image Classifier dataset (natural scene classification). PyTorch resizing, normalization and tensor conversion were used to preprocess all datasets. Multi-class classification tasks of the ResNet18 with pretrained ImageNet weights were fine-tuned. The trained models were provided through a low-code FastAPI interface, which was image based using Docker, and deployed on Kubernetes. Jenkins and ArgoCD CI\/CD pipelines based on automated version control, testing, deployment and continuous monitoring. Some of the evaluation measures were accuracy, precision, recall, F1-score, and the confusion matrix analysis. Results & Comparison: ResNet18 demonstrated the highest accuracy (93.27%), precision (93.31), recall (93.27), and F1-score (93.26), outperforming YOLOv5s (76.3%), LSTM, Random Forest (86 90s), and other machine learning models. Visualization ensured correct detection of every type of defect. Conclusion: CI\/CD pipelines that are combined with Docker, Kubernetes, and low-code AI development will allow having reproducible, scalable, and enterprise-ready deployment. The framework is operational efficient, adaptable retraining and limited manual intervention and is a feasible solution to real time industrial AI use.