基于超分重建与Dy-YOLOv7的茶叶分级识别

Classification Identification of Tea Based on SuperfractionReconstruction and Dy-YOLOv7

目的通过对优质茶叶嫩芽不同姿态的识别,为一体化的选择性采收提供技术支持。方法通过Real-ESRGan网络对部分茶芽图像进行重建,丰富图像中包含的特征信息;后使用Dy-YOLOv7网络检测茶叶嫩芽,首先,在Backbone和Neck侧分别嵌入SE、CBAM和EMA 3种注意力机制,探索注意力机制的最佳嵌入点,最终构建SE-ELAN-H模块,提升层内特征的提取能力;其次,将检测头部IDetect_Head替换为具有统一尺度感知、空间感知和任务感知的DyHead(Dynamic Head),以获得更强的特征表达能力;最后,使用MPDIoU(Maximum Partially Differentiable IoU)作为模型边界框损失函数,克服IoU损失函数的不可导性以及局部最优问题,让模型预测边界框更聚焦于嫩芽目标。结果Dy-YOLOv7算法对于茶叶单芽的平均精度均值为91.6%,一芽一叶为92.5%,一芽二叶为94.6%,与原始YOLOv7相比,精度分别提高了5.5%、2.1%和3.8%。结论该方法在兼顾准确率的前提下,实现了对优质茶苗不同姿态的识别,可为茶叶嫩芽分级识别、智能化采摘提供重要的理论基础。

Objective To provide technical support for integrated selective harvesting by identifying the different posture of high quality tea buds.Methods Real-ESRGan network was used to reconstruct part of tea bud image to enrich the feature information contained in the image.After that,Dy-YOLOv7 network s used to detect the tea bud.Firstly,SE,CBAM and EMA re embedded in the Backbone and Neck side respectively to explore the best embedding point of the attention mechanism construct to improve the extraction ability of intra-layer features.Secondly,IDetect_Head s replaced by DyHead(Dynamic Head),which ha unified scale perception,space perception and task perception,so as to obtain stronger feature representation ability.Finally,MPDIoU(Maximum Partially Differentiable IoU)s used as the model boundary frame loss function to overcome the non-differentiability of IoU loss function and local optimization,so that the model prediction boundary frame s more focused on the bud target.Results The average accuracy of Dy-YOLOv7 algorithm for single bud of tea was 91.6%,92.5%for one bud and one leaf,94.6%for one bud and two leaves ompared with original YOLOv7,was improved by 5.5%,2.1%and 3.8%,respectively.Conclusion This method realize the recognition of different posture of high-quality tea seedlings,provid an important theoretical basis for the classification of tea buds and intelligent picking.