基于改进AttentionGAN与YOLO v8n的食用玫瑰夜间成熟度轻量化识别方法

Lightweight Method for Nighttime Maturity Recognition of Edible Roses Based on Improved AttentionGAN and YOLO v8n

摘要:

为解决云南食用玫瑰夜间自动化采摘中样本稀缺、成熟度识别精度低、小目标漏检及边缘设备算力受限等问题,本研究设计了改进AttentionGAN和YOLO v8n方法。在数据增强环节,构建基于改进AttentionGAN的日夜图像转换模型,引入多尺度注意力机制(Multi-scaleattention)、设计渐进式融合策略,同步对生成器进行轻量化改进;在检测模型环节,以YOLO v8n为基础构建YOLO-NRP模型,采用RVG-Ghost(RepVGG-Ghost)模块替换C2f主干、在主干网络末端嵌入ECA-Spatial注意力机制以增强关键特征表达、采用PAFPN-Lite优化颈部网络,强化多尺度特征融合效果。试验结果显示,改进AttentionGAN生成图像的FID(Frechet inception distance)降至87.97、SSIM(Structural similarity index)提升至0.5882,质量显著优于原模型,有效扩充了夜间数据集;YOLO-NRP模型的精确率、召回率、mAP50和mAP50-95分别为:91.3%、90.2%、94.3%、80.7%,所有指标均优于YOLO v8n基线模型,同时模型内存占用量仅5.49MB,推理速度达70.26f/s,满足边缘设备部署要求,可为自动化采摘装备研发提供可靠技术支撑。

Abstract:

Aiming to address the challenges in nighttime automated harvesting of edible roses in Yunnan, including limited samples, low maturity recognition accuracy, missed small targets, and constrained edge computing capacity, the targeted technical optimizations were conducted. In the data enhancement stage, an improved AttentionGAN-based day-night image conversion model was constructed, incorporating the multi-scale attention mechanism, designing a progressive fusion strategy, and simultaneously implementing lightweight modification of the generator. For the detection model, the YOLO-NRP model was built based on YOLO v8n: the C2f backbone was replaced with the RVG-Ghost (RepVGG-Ghost) module, the ECA-Spatial attention mechanism was embedded at the end of the backbone network to enhance key feature expression, and the neck network was optimized with PAFPN-Lite to strengthen multi-scale feature fusion. Experimental results showed that the images generated by the improved AttentionGAN achieved a Frechet inception distance (FID) of 87. 97 and a structural similarity index (SSIM) of 0. 588 2, with significantly better quality than that of the original model, effectively expanding the nighttime dataset. The YOLO-NRP model achieved 91. 3% precision, 90. 2% recall, 94. 3% mAP50, and 80. 7% mAP50 95, with all indicators outperforming the YOLO v8n baseline model. Meanwhile, the model weighed only 5. 49 MB and had an inference speed of 70. 26 f/ s, meeting the requirements of edge device deployment and providing reliable technical support for the development of automated harvesting equipment.