<p>Automated defect detection in manufacturing has gained significant attention. In casting manufacturing, quality inspection is critical to ensure product reliability and reduce waste, yet manual inspection methods currently in use remain ineffective. While computer vision systems show promise, most studies evaluate models on curated datasets without considering deployment-specific constraints. Comparative evaluations of multiple architectures for domain-specific defect detection remain limited, leaving a gap in practical guidance for model selection in real-world quality control systems. This study addresses this gap by evaluating two convolutional neural network (CNN) architectures, a custom CNN and MobileNetV2, for detecting defects in cast submersible pumps’ impellers. Both models were assessed using accuracy, precision, recall, F1-score, and inference speed, providing insights into their suitability for deployment. This study contributes to the literature by comparing a lightweight pretrained model (MobileNetV2) and a custom CNN for industrial casting defect detection, linking model choice to deployment feasibility. The results showed high performance for both models, with the custom CNN achieving faster inference (0.2532 s/image) compared to MobileNetV2 (0.3107 s/image). In an industrial setting where thousands of components may be inspected daily, this speed difference can translate into minutes or hours saved per production cycle. Statistical tests revealed no significant difference in accuracy, confirming both models as viable solutions. In real-world deployments, challenges like variability in conditions may affect detection accuracy, requiring data augmentation and periodic retraining; thus, our findings extend prior research by shifting focus from pure accuracy metrics toward deployment-oriented evaluation, providing actionable guidance for industrial AI inspection systems.</p>