Synthetic diamonds are prone to subtle defects during the manufacturing process, which can significantly impact their quality grading and market value. Due to the highly symmetrical and complex geometric structures of synthetic diamonds, accurately detecting defects of varying scales in complex backgrounds poses a considerable challenge. To address this issue, we propose a target detection algorithm specifically designed for synthetic diamond quality evaluation, named Diamond-DETR, aiming to improve detection accuracy and model generalization under resource-constrained environments. Diamond-DETR optimizes the backbone network by introducing a lightweight multi-scale feature extraction module and utilizes the RepFasterNet block to reduce computational complexity and enhance inference speed. Additionally, it incorporates an encoder structure based on a high-level screening-feature fusion pyramid network, which employs channel attention mechanisms to filter and fuse features at different scales, thereby enhancing the model’s ability to detect defects of various sizes. Furthermore, a cross-stage fusion module is introduced, which leverages dilated convolutions to expand the receptive field without increasing computational cost, improving the model’s capacity to perceive long-range dependencies and complex geometric structures. Experimental results demonstrate that Diamond-DETR outperforms the original RT-DETR model in terms of parameter efficiency, inference speed, and detection accuracy, making it particularly well-suited for deployment in resource-constrained inspection scenarios. The model also exhibits competitive performance in a cross-dataset evaluation on an external industrial dataset, indicating potential applicability in related industrial inspection scenarios.
Yan et al. (Sun,) studied this question.