Strawberry ripeness detection in complex environment based on improved RT-DETR

Abstract

Accurate and rapid detection of strawberry (Fragaria × ananassa Duchesne ex Rozier) maturity in greenhouse environments is critical for advancing mechanized harvesting, yet existing methods struggle with challenges such as small target sizes, dense clustering, and occlusion by foliage. The real-time detection transformer (RT-DETR), as a real-time end-to-end detector, eliminates the need for NMS processing and provides a baseline for real-time detection. But its performance is limited by computational inefficiency and insufficient robustness in complex agricultural scenarios. To address these limitations, we propose an enhanced strawberry maturity detection model, partical ghost convolution deformable attention simple parameter free and efficient local high feature fusion detection transformer (PDSE-DETR). The backbone network is enhanced using lightweight modules to reduce model complexity while feature extraction capability is maintained. Integrating attention mechanisms with feature pyramids to minimize background interference, boosting detection of densely clustered targets. Optimizing the loss function to improve localization accuracy for small target regression. The PDSE-DETR was validated using the strawberry dataset created in this study. Experimental results demonstrate that PDSE-DETR achieves a 2.1% improvement in average detection accuracy over RT-DETR, while reducing parameters and computational costs by 30.2% and 30.7%, respectively. These advancements enable reliable real-time maturity assessment in practical greenhouse environments, offering a scalable solution to optimize automated strawberry harvesting and reduce operational inefficiencies.