基于RT-DETR改进的幼桃识别模型OA

In response to the challenges of identifying immature peaches in natural environments,such as color similarity with surrounding environments,uneven lighting,and obstruction by branches and leaves,a detection model FREDC-RTDETR was proposed based on improving RT-DETR-R18 in this study.By replacing the BasicBlock in the RT-DETR-R18 backbone network with the Faster NetBlock,incorporating RepConv reparameterization technology,and introducing the EMA attention mechanism,a new backbone network FRE Block was designed,which could reduce the number of parameters but enhancing the model's feature extrac-tion capability.In the neck network,the original AIFI module was replaced with AIFI-LPE based on learnable position encoding to address the issue of attention shift,and the DySample dynamic upsampling along with the redesigned CG block Down downsampling operator were employed to optimize the upsampling and downsam-pling processes.Additionally,the Shape-IoU loss function was used to enhance the model's ability to capture image details.The experimental results showed that on the self-built dataset,the improved model achieved the mean average precision of 96.1％,the recall rate of 91.9％,and the precision of 97.6％,representing increa-ses of 2.4,2.7 and 2.5 percentage points compared to the original model,respectively.In conclusion,the pro-posed model in this study demonstrated better robustness and accuracy in complex backgrounds,which could provide a reference for early yield prediction and green fruit identification of fruit trees.