基于改进YOLO v7-tiny的玉米种质资源雄穗检测方法

Tassel Detection Method of Maize Germplasm Resources Based on Improved YOLO v7-tiny

摘要:

针对玉米种质资源遗传多样性丰富导致雄穗大小、形态结构及颜色呈现较大差异，无人机搭载可见光传感器相比地面采集图像分辨率低，以及图像中部分雄穗过小、与背景相似度高、被遮挡、相互交错等情况带来的雄穗检测精度低的问题，提出了一种改进YOLO v7-tiny模型的玉米种质资源雄穗检测方法。该方法通过在YOLO v7-tiny中引入SPD-Conv模块和VanillaBlock模块，以及添加ECA-Net模块的方式，增强模型对雄穗特征的提取能力。利用自建的玉米种质资源雄穗数据集，训练并测试改进模型。结果表明，改进YOLO v7-tiny的平均精度均值为94.6%，相比YOLO v7-tiny提升1.5个百分点，相比同等规模的轻量级模型YOLO v5s、YOLO v8s分别提升1.0、3.1个百分点，显著降低了图像中雄穗漏检及背景误检为雄穗的发生，有效减少了单穗误检为多穗和交错状态下雄穗个数误判的情况。改进YOLO v7-tiny模型内存占用量为17.8MB，推理速度为231f/s。本文方法在保证模型轻量化的前提下提升了雄穗检测精度，为玉米种质资源雄穗实时、精准检测提供了技术支撑。

关键词:种质资源;玉米雄穗;目标检测;无人机;可见光影像;YOLO v7-tiny

Abstract:

Due to the rich genetic diversity of maize germplasm resources, the size, morphological structure and color of tassels were quite different. The resolution of maize tassel image collected by UAV equipped with visible light sensor was lower than that of ground acquisition, and some tassels in the image were too small, which were highly similar to the background, occluded and interlaced. The above factors led to low accuracy of tassel detection. Therefore, a tassel detection method for maize germplasm resources based on improved YOLO v7-tiny model was proposed. This method enhanced the model’s ability to extract tassel features by introducing SPD-Conv module and VanillaBlock module into YOLO v7-tiny, and adding ECA-Net module. Tested on the self-built tassel dataset of maize germplasm resources, the mean average precision of the improved YOLO v7-tiny was 94.6%, which was 1.5 percentage points higher than that of YOLO v7-tiny, and 1.0 percentage points and 3.1 percentage points higher than that of the lightweight models YOLO v5s and YOLO v8s, respectively. This method significantly reduced the occurrence of missing tassels and false detection of background as tassels in the image, and effectively reduced the misdetection of a single tassel as multiple tassels and the number of tassels in interlaced state. The model size of the improved YOLO v7-tiny was 17.8MB, and the inference speed was 231f/s. The proposed method can improve the accuracy of tassel detection under the premise of ensuring the lightweight of the model, and can provide technical support for the real-time and accurate detection of tassel of maize germplasm resources.

Key words:germplasm resources;maize tassel;target detection;UAV;visible light image;YOLO v7-tiny