基于CenterNet的自然场景下苹果快速识别方法  

CenterNet-based approach for fast apple recognition method in natural scenes

作　　者：樊攀 孙瑾 周桥 陈曌宇 FAN Pan;SUN Jin;ZHOU Qiao;CHEN Zhao-yu(School of Computer,Baoji University of Arts and Sciences,Baoji 721016,Shaanxi,China)

机构地区：[1]宝鸡文理学院计算机学院,陕西宝鸡721016

出　　处：《宝鸡文理学院学报（自然科学版）》2024年第3期56-63,共8页Journal of Baoji University of Arts and Sciences(Natural Science Edition)

基　　金：陕西省教育厅青年创新团队科研计划项目(23JP004)。

摘　　要：目的构建一种高效的苹果目标识别方法,提升苹果采摘机器人在果园自然场景中的目标识别准确率和效率。方法通过采用CenterNet神经网络为检测框架,同时融入了分组卷积和深度可分离卷积的理念,设计了一种基于瓶颈结构堆叠策略的轻量级特征提取网络Light-Weight Net。结果设计了一种适配于苹果采摘机器人视觉系统的识别算法,实现了在果园自然场景中高精度和高效率的苹果目标识别。结论该模型在测试集上实现了96.60%的目标识别准确率(以平均精度衡量),通过与YOLOv3和Efficient-D0模型在相同测试集进行对比,试验结果平均精度分别提高了6.30%和5.17%,单幅图像平均识别时间分别快了0.014 s和0.05 s。Purposes—To design a fast apple target recognition method for improving the target recognition accuracy and recognition efficiency of an apple picking robot in a natural scene in an orchard.Methods—The CenterNet neural network is used as the detection framework,and a Light-Weight Net lightweight feature extraction network is proposed by drawing on the ideas of grouped convolution and depth-separable convolution.Results—The recognition algorithm adapted to the visual system of apple picking robots has been designed which achieves high-precision and efficient apple target recognition in natural orchard scenes.Conclusions—The model recognized an AP value of 96.60%under the test set,and by comparing with YOLOv3 and Efficient-D0 model in the same test set,the experimental results showed that the AP value was improved by 6.30%and 5.17%,and the average recognition time of a single image was faster by 0.014 s and 0.05 s.

关 键 词：采摘机器人 自然场景 苹果识别 CenterNet 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]