Litchi flower targets integrating multi-scale attention and decoupled detection

In commercial litchi production, the scientific regulation of floral load is critical for ensuring consistent yields and high-quality fruits. However, litchi exhibits profuse and concentrated flowering. Accurately detecting these dense floral clusters in complex orchard environments is highly challenging. To address this challenge, this study proposes the Fusion Decoupled and Multi-scale Attention Network (FDMA-Net), an improved lightweight object detection model. The model employs High Performance GPU Network V2 (HGNetV2) as the backbone, incorporating depthwise separable convolutions (DWConv) to create a strong yet lightweight feature-extraction framework. A parameter-free simple attention module (SimAM) is integrated at the backbone output to enhance fine-grained feature discrimination. The C3k2_Star module, an enhanced cross-stage feature fusion block, reinforces context awareness and multi-scale feature fusion. Finally, a decoupled detection head integrating the spatial enhancement attention module (Detect_SEAM) leverages task decoupling and feature enhancement to improve localization accuracy and confidence in dense scenes. Experimental results indicate that the proposed model achieves a mean Average Precision (mAP@0.5) of 94.1%, outperforming the YOLOv11n baseline by 4.2%. Simultaneously, it reduces both the model size and floating-point operations (FLOPs) by 18.9% and 17.5%, respectively. Furthermore, robustness evaluations confirm that the model maintains high stability under non-ideal conditions. Consequently, the proposed method delivers superior performance in litchi flower detection, making it highly suitable for real-time, high-precision deployment on resource-constrained edge devices.