Empty Glass Bottle Defect Detection Based on Deep Learning with CNN Using SSD MobileNetV2 Model

Reusing glass bottles is one of the many ways to help reduce pollution and waste, and detecting defects is an important part of the glass bottle reusing sector to prevent damaged products from reaching their end customers. As a result, costs from withdrawals or recalls are reduced, resources are freed up for other areas of the line, productivity is increased, and quality, reliability, and safety are assured, which improves the company’s reputation and brand loyalty. But the majority of micro, small, and medium enterprises (MSMEs) that use clear glass bottles as their reused products rely only on manual inspection. These are vulnerable to human error, resulting in poor product quality and rejections. To overcome these problems, this research presents a cost-effective deep learning-based method using the SSD MobileNetV2 model for detecting generic glass bottle defects in an enclosed space environment that are macro and micro in size. These defects are found in the body and base of the bottle and are caused by impact or dirt, such as micro-cracks, dirty wares, bruises, checks, chips, and broken wares. With the use of transfer learning and data augmentation, the results revealed that the device is accurate at detecting glass bottle defects, with up to 98.07% overall system accuracy. The findings of the study provided an insight into defect detection in real-world glass bottle reusing settings.