Deep Learning-Powered Visual Inspection Using SSD Mobile Net V1 with FPN

There is an increasing demand to automate manufacturing inspection processes that can be time-consuming and subject to the expertise of individual inspectors. While there are advantages to having humans assess the quality of a part versus automated vision systems, such as greater flexibility of the inspector to adapt to new parts, this approach is also prone to errors. This study examines the capability of optical inspection techniques to reduce inspection errors in remanufacturing. We implemented an SSD Mobile-Net Algorithm that uses depth-wise separable convolutions to build lightweight deep neural networks and Feature Pyramid Network (FPN) to enhance feature extraction. The algorithm was evaluated using the GC10-DET benchmark data. The algorithm addressed the extreme imbalance between common and uncommon defect samples. The algorithm’s performance is compared with other object detection algorithms using average precision and mean average precision (mAP) metrics. Our model outperforms five class categories compared to other state-of-the-art models, especially in small size defect and class imbalance categories. The importance of data pre-processing is also discussed, including improving data quality while keeping the training model constant.