Automated visual inspection plays a central role in electronics manufacturing processes involving UV adhesive deposition, where undetected defects may compromise mechanical stability and lead to latent failures. In NVMe solid-state drive (SSD) production, inspection systems must prioritize reliability and defect containment, as false negatives represent a critical operational risk. Despite this requirement, many deep learning-based inspection approaches remain optimized for global accuracy metrics, which are not fully aligned with industrial reliability constraints. This paper presents a convolutional neural network (CNN)-based visual inspection framework tailored for high-reliability deployment in NVMe SSD manufacturing. The proposed approach emphasizes inspection problem formulation rather than architectural complexity. A physically coherent region of interest (ROI) was defined to encompass the functional UV adhesive deposition area surrounding the SSD controller, reducing background interference. Three CNN backbones—ResNet50, EfficientNetV2, and MobileNetV2—were evaluated under identical conditions using transfer learning. Additionally, multiple decision strategies, including calibrated decision thresholds, were analyzed to reduce false negatives while controlling false-positive rates. All experiments were conducted using an industrial dataset collected from an operational production line. Experimental results indicate that ResNet50 achieved stable accuracy at around 86% but showed limited reliability due to elevated false-positive rates. In contrast, EfficientNetV2 and MobileNetV2 achieved a defective part recall above 98%, overall accuracy exceeding 91%, and a reduction of more than 50% in false negatives compared to the ResNet50 baseline. MobileNetV2 matched the performance of EfficientNetV2 while maintaining lower computational complexity. The results demonstrate that reliability gains in industrial visual inspection are more effectively achieved through region-focused analysis and decision-level optimization than by increasing model complexity. The proposed framework is suitable for deployment in high-throughput manufacturing environments and provides a basis for future extensions involving hybrid inspection strategies.