An improved RT-DETR approach for real-time detection of densely occluded blueberries and micro-calyxes

Accurate detection of blueberries and their micro-scale calyx regions is important for visual perception in robotic blueberry harvesting, especially in densely clustered scenes where fruits are frequently occluded and adjacent to each other. However, blueberry calyx detection remains challenging because the front calyx occupies only a small image region and is easily affected by background interference, illumination variation, and feature attenuation during deep feature extraction. To address these challenges, this study proposes MFE-DETR, an improved RT-DETR-based real-time detection model for densely occluded blueberries and micro-calyxes. A Micro-Feature Enhanced Attention (MFEA) module is introduced to strengthen the representation of weak calyx features by combining heterogeneous local receptive fields with a spatial prior bias. The proposed module is designed to enhance the association between micro-calyx regions and the surrounding fruit context while maintaining the end-to-end detection framework of RT-DETR. Experimental results on a self-constructed blueberry image dataset show that MFE-DETR achieved an overall [email protected] of 94.2 %, with a Front calyx [email protected] of 91.5 %. In addition, deployment experiments on an NVIDIA Jetson AGX Orin platform showed that the proposed model achieved real-time edge-side inference under the tested configuration, suggesting its potential as a visual perception module for subsequent robotic blueberry harvesting systems.