Improved RT-DETR and its application to fruit ripeness detection

Yuhao Cao ¹ Bai Yun ^2*

• ¹ Wuhan University of Science and Technology, Wuhan, Hubei Province, China
• ² National University of Defense Technology, Changsha, China

Crop maturity status recognition is a pivotal component of automated harvesting. Traditional manual detection methods are plagued by inefficiency and high costs. To address this challenge, we propose enhancements to the Real-Time DEtection TRansformer（RT-DETR） method for crop maturity detection. We refine the original model's Backbone structure: firstly, we enhance the HG Block ( Highway Gated Block) by substituting the conventional convolution with the Rep Block during feature extraction. The incorporation of multiple branches in this block proves beneficial for enhancing model accuracy. Subsequently, we replace the conventional convolution in the Rep Block with Partial Convolution (PConv). PConv applies convolution solely to a portion of the input channel system, thus mitigating computational redundancy. Lastly, we introduce the Efficient Multi-Scale Attention （EMA）mechanism and adopt a cross-space learning approach to ensure uniform distribution of spatial semantic features within each feature group, thereby yielding superior results in terms of model parameters. We conduct ablation experiments and juxtapose the RT-DETR model with its familial counterparts to substantiate the accuracy and efficacy of our proposed enhancements. The experimental findings affirm the efficacy of the refined algorithm in discerning crop maturity levels, resulting not only in improved detection accuracy but also in reductions in model parameters and computational complexity. Relative to the original model, the average accuracy mAP@0.5 witnessed a 2.9% enhancement, with a reduction in model size by 5.5% and computational complexity by 9.6%. In order to verify the generality of the model, we also compared the standard RT-DETR model, the YOLOv8 model, and the improved model on the plant pest detection datasets, and the experimental results show that the improved model outperforms the other models for general scenarios. This method holds significant implications for the advancement of automated crop maturity detection methodologies.