基于频域数据增强与轻量化YOLO v7模型的成熟期香梨目标检测方法

Mature Stage Pear Detection Method Based on Frequency Domain Data Augmentation and Lightweight YOLO v7 Model

为实现香梨自动化采摘,本文以YOLO v7-S为基础模型,针对果园中香梨果实、果叶和枝干之间相互遮挡,不易精准检测的问题,设计了一种轻量化香梨目标检测M-YOLO v7-SCSN+F模型。该模型采用MobileNetv3作为骨干特征提取网络,引入协同注意力机制(Coordinate attention,CA)模块,将YOLO v7-S中的损失函数CIoU替换为SIoU,并联合Normalized Wasserstein distance(NWD)小目标检测机制,以增强网络特征表达能力和检测精度。基于傅里叶变换(Fourier transform,FT)的数据增强方法,通过分析图像频域信息和重建图像振幅分量生成新的图像数据,从而提高模型泛化能力。实验结果表明,改进的M-YOLO v7-SCSN+F模型在验证集上的平均精度均值(mAP)、精确率和召回率分别达到97.23%、97.63%和93.66%,检测速度为69.39 f/s,与Faster R-CNN、SSD、YOLO v3、YOLO v4、YOLO v5s、YOLO v7-S、YOLO v8n、RT-DETR-R50模型在验证集上进行性能比较,其平均精度均值(mAP)分别提高14.50、26.58、3.88、2.40、1.58、0.16、0.07、0.86个百分点。此外,改进的M-YOLO v7-SCSN+F模型内存占用量与YOLO v8n和RT-DETR-R50检测模型对比减少16.47、13.30 MB。本文提出的检测模型对成熟期香梨具有很好的目标检测效果,为背景颜色相近小目标检测提供参考,可为香梨自动化采摘提供有效的技术支持。

In the practice of modern agricultural production,the method of harvesting agricultural products is gradually shifting towards mechanization and intelligence.An increasing number of robots are being introduced into actual production and progressively replacing traditional manual labor.However,in natural environments,factors such as weather,lighting,the similarity in color between fruits and their backgrounds,and mutual occlusion between fruits and branches significantly increased the difficulty of fruit target detection.To accurately detect pears in natural environments,a lightweight pear detection method M YOLO v7 SCSN+F was designed based on the YOLO v7 S foundational model.This model introduced MobileNetv3 into the YOLO v7 S model as its backbone feature extraction network,thereby reducing the number of parameters in the network.It incorporated a coordinate attention(CA)mechanism in the model's feature fusion layer to enhance the network's feature representation capabilities.The loss function CIoU in YOLO v7 S was replaced with SIoU,which was used in conjunction with the normalized Wasserstein distance(NWD)mechanism for small target detection,further improving the detection accuracy for fragrant pears.Based on the Fourier transform(FT)data augmentation method,new image data was generated by analyzing the frequency domain information of images and reconstructing the amplitude components,thereby enhancing the model's generalization ability.Experimental results showed that the improved M YOLO v7 SCSN+F model achieved mean average precision(mAP),precision,and recall rates of 97.23%,97.63% and 93.66%,respectively,on the validation set,with a detection speed of 69.39 f/s.The proposed detection model improved performance compared with Faster R CNN,SSD,YOLO v3,YOLO v4,YOLO v5s,YOLO v7 S,YOLO v8n and RT DETR R50 models on the validation set,with mean average precision(mAP)enhancements of 14.50,26.58,3.88,2.40,1.58,0.16,0.07 and 0.86 percentage points,respectively.Furthermore,the improved M YOLO v7 SCSN+F model reduced its parameter count by 16.47 MB and 13.30 MB,respectively,when compared with the advanced YOLO v8n and RT DETR R50 detection models.The detection model introduced demonstrated a high degree of effectiveness in target detection for mature pears,offering a reference for detecting small objects with backgrounds of similar color,and provided effective technical support for the automation of pear harvesting.