Authors

Abstract

To achieve stable abnormal egg detection under edge device deployment conditions, this paper proposes a CSDE-DETR-R18 abnormal egg detection model based on an optimized RT-DETR-R18. The CSFM module is incorporated into the feature fusion path of the RT-DETR-R18 model to aggregate multi-scale features. Subsequently, the model's ability to extract low-contrast features is enhanced by replacing the standard convolutions in the backbone network with the DEConv module, which improves its recognition capability for abnormal egg characteristics such as light-colored spots. Finally, the NWD+AIOU mixed loss function is employed to improve the model's localization accuracy for minute targets (such as fine cracks, specks, etc.). The experimental results demonstrate that the CSDE-DETR-R18 model achieved mAP@0.5, P, and R of 87.8%, 88.9%, and 87.3%, respectively, representing improvements of 3.4, 3.1, and 0.5 percentage points over RT-DETR-R18. Test results from the Raspberry Pi deployment revealed that CSDE-DETR-R18 demonstrated a significant advantage in inference stability compared with the YOLOv8 series, one of the most advanced detectors currently available. This is of great engineering significance for tasks requiring stable timing performance at equal heights, such as robotic arm sorting.

Abstract in Chinese

为在边缘设备部署条件下实现稳定的异常鸡蛋识别，本文提出一种基于改进RT-DETR-R18的异常鸡蛋检测模型CSDE-DETR-R18。在RT-DETR-R18的进步特征融合路径加入CSFM模块聚合多尺度特征；然后使用DEConv模块替换骨干网络中的标准卷积提高模型的低对比度特征提取能力，增强对浅色斑点等异常鸡蛋特征的识别能力；最后用NWD+AIOU混合损失函数来提升模型对于微小目标（细小裂纹、斑点等）的定位精度。实验结果显示，使用CSDE-DETR-R18模型得到的mAP@0.5、P和R分别为87.8%、88.9%和87.3%，比RT-DETR-R18提升了3.4、3.1和0.5个百分点。树莓派部署测试结果显示，相较于目前最先进检测器之一的YOLOv8系列，CSDE-DETR-R18具有明显的推理稳定性优势。这对机器臂分拣等高时序稳定性要求的任务具有重大的工程意义。