Dataset Purification-Driven Lightweight Deep Learning Model Construction for Empty-Dish Recycling Robot

To solve the labor shortage, robots have dramatically changed the world by combining powerful deep learning (DL) technology. Certainly, DL technology has become the key point of the widespread robot application. Efficient DL models depend on high-quality datasets and optimized architectures. However, some open datasets contain anomalous data that degrade model performance. Moreover, complex structures and high computational costs limit the adoption of DL models. This study proposes a dataset purification-based lightweight DL model construction strategy to solve these challenges. Initially, a dataset purification method is developed to filter out the anomaly data in a newly created dataset, which utilizes a lightweight cross-scale DL model OGNet to detect the anomaly data to achieve dataset purification. Subsequently, a highly efficient lightweight OGNet-based object detection (OD) model family, YOLO-OG, is presented to train the purified dataset. To evaluate the proposal, the strategy is implemented on the Empty-dish Recycling Robot. Experiments show that OGNet achieves excellent accuracy with only 0.68 M parameters and 0.35 GFLOPs. On purification Dish-10 dataset, the mean Average Precision (mAP) of YOLO-OG increases a maximum of 4.28% than original Dish-10 dataset. Meanwhile, YOLO-OG outperforms other advanced OD models, achieving the best accuracy of 99.20% mAP and the smallest 1.60 M parameters. YOLO-OG also reaches 99.86% mAP on the Dish-20 open dataset. On three other open datasets, YOLO-OG also shows excellent performance and surpasses most of the other OD models, which confirms the strong generalization ability of YOLO-OG.