基于改进YOLO 12的复杂环境红花轻量化识别方法

Lightweight Safflower Recognition Method Based on Improved YOLO 12 for Complex Environments

摘要:

针对田间光照变化和枝叶花丝重叠遮挡等复杂背景干扰下红花识别发生漏检、识别准确度低及模型体积过大不利于边缘设备部署等问题,提出了一种基于YOLO 12改进的轻量化网络结构LNSR(Lightweight network for safflower recognition)。首先设计了HEPDSF(Heterogeneous edge-pooling dual-stream fusion module)异构边缘池化双流融合模块,增强网络在初始阶段对红花边缘和纹理信息的提取能力,提高识别准确度;然后构建了TriCAFusion(Triple cooperative adaptive fusion module)三重协同自适应融合模块,增强模型对目标关键特征表征能力,降低漏检概率;再次引入AdaPool(Adaptive neighborhood pooling for down sampling module)自适应邻域池化下采样模块,增强模型对光照变化和遮挡场景的鲁棒性;最后开发LSBNet(Lightweight shared-BN network)轻量化共享卷积分离BN网络检测头,降低模型复杂度,提高模型在边缘设备部署能力。试验结果表明,LNSR在红花数据集上参数量仅1.72×106,较YOLO 12降低31.5%,模型内存占用量为4.4MB,较YOLO 12降低1.1MB,mAP50达98.9%,召回率98.3%,较YOLO 12分别提升1.4、1.0个百分点。将LNSR泛化至菊花数据集,参数量仅1.63×106,模型内存占用量为4.4MB,mAP50达97.0%,较YOLO 12提升5.9个百分点,通过热力图验证其对花瓣边缘与花蕊纹理的精准聚焦和表征能力,部署于边缘设备Jetson实时帧率达到30f/s。结果表明,LNSR通过四模块协同创新,实现精度、轻量化较优平衡,为红花选择性采收方式提供了高效可靠的视觉识别方案。

Abstract:

Aiming to address the challenges of missed detection, low recognition accuracy caused by complex field conditions such as varying lighting and occlusion by overlapping branches and filaments, as well as the difficulty in deploying large models on edge devices, lightweight network for safflower recognition ( LNSR), an improved lightweight network structure was proposed based on YOLO 12. Specifically, the heterogeneous edge-pooling dual-stream fusion module (HEP-DSF) was designed to enhance the network’s ability to extract edge and texture information in the early stage, thereby improving recognition accuracy. The triple cooperative adaptive fusion module (TriCAFusion) was constructed to strengthen the model??s representation of key target features and reduce the probability of missed detection. The adaptive neighborhood pooling for down sampling module (AdaPool) was introduced to improve the model’s robustness to lighting variations and occluded scenes. Furthermore, the lightweight shared-BN network (LSBNet) detection head was developed to reduce model complexity and improve deployment efficiency. Experimental results on the safflower dataset showed that LNSR achieved only 1. 72 × 106 parameters, a reduction of 31. 5% compared with that of YOLO 12, with a model size of 4. 4 MB (1. 1 MB smaller). It also reached an mAP50 of 98. 9% and a recall rate of 98. 3%, representing improvements of 1. 4 and 1. 0 percentage points, respectively. When generalized to the chrysanthemum dataset, LNSR achieved 1. 63 × 106 parameters, a model size of 4. 4 MB, and an mAP50 of 97. 0%, which was 5. 9 percentage points higher than that of YOLO 12. Heatmap validation confirmed its precise focus and characterization ability on petal edges and stamen textures. Deployed on the Jetson edge device, it achieved a real-time frame rate of 30 f/ s. It demonstrated that through the collaborative innovation of the four modules, LNSR achieved an optimal balance between accuracy and lightweight design, providing an efficient and reliable visual recognition solution for selective harvesting of safflower.