<![CDATA[This study addresses the need for affordable visual inspection support in micro, small, and medium enterprises (MSMEs) engaged in screen-printing production. Although machine vision and deep learning have been widely applied in manufacturing quality control, many existing systems are designed for relatively controlled industrial settings and require stable cameras, lighting, computing resources, and technical expertise. This condition limits direct adoption by small MSMEs, where image acquisition is often performed with operator-level devices under variable lighting and background conditions. This study designed and evaluated an initial low-resource visual inspection pipeline consisting of low-cost image acquisition, five-class defect labeling, MobileNetV3-based transfer learning, performance evaluation, and TensorFlow Lite conversion. The dataset consisted of 160 screen-printing images grouped into five classes: good, misalignment, bleeding, pinholes, and ghosting. The preliminary evaluation yielded 24.38% multiclass accuracy and a loss of 2.5635, indicating that the model could not yet reliably distinguish detailed defect categories. The converted TensorFlow Lite model was 5.43 MB, indicating that the technical conversion path was feasible. A binary quality-control interpretation produced 75.63% accuracy, but 27 defective images were still predicted as pass QC. Therefore, the pipeline cannot be treated as a final quality-control decision system. The main contribution of this study is empirical evidence that image-acquisition quality, dataset sufficiency, class separability, and training configuration are critical bottlenecks in developing lightweight deep-learning-based inspection for low-resource MSME environments.]]>
Copyrights © 2026
