欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2022, Vol. 43 ›› Issue (4): 580-589.DOI: 10.11996/JG.j.2095-302X.2022040580

• 图像处理与计算机视觉 • 上一篇    下一篇

基于多尺度特征实现超参进化的野生菌分类研究与应用

  

  1. 1. 南昌工程学院信息工程学院，江西 南昌 330000；
    2. 南昌工程学院机械工程学院，江西 南昌 330000
  • 出版日期:2022-08-31 发布日期:2022-08-15
  • 通讯作者: 黄志开(1969)，男，教授，博士。主要研究方向为图像处理等
  • 作者简介:张盾(1996)，男，硕士研究生。主要研究方向为图像处理与目标检测
  • 基金资助:
    国家重点研发计划项目(2019YFB1704502)；国家自然科学基金项目(61472173)；江西省教委资助项目(GJJ151134)

Research and application of wild mushrooms classification based on multi-scale features to realize hyperparameter evolution

  1. 1. School of Information Engineering, Nanchang Institute of Technology, Nanchang Jiangxi 330000, China;
    2. School of Mechanical Engineering, Nanchang Institute of Technology, Nanchang Jiangxi 330000, China
  • Online:2022-08-31 Published:2022-08-15
  • Contact: HUANG Zhi-kai (1969), professor, Ph.D. His main research interest covers image processing, etc
  • About author:ZHANG Dun (1996), master student. His main research interests cover image processing and target detection
  • Supported by:
    National Key Research and Development Plan of China (2019YFB1704502); National Natural Science Foundation of China (61472173);
    The Grants from the Educational Commission of Jiangxi Province of China (GJJ151134)

PDF (PC)

141

摘要：

在我国，因误食不可食用野生菌而导致中毒的事件频发，尤其是云南等西南地区，由于野生菌种类的类间特征差异较小，且实际场景下的图像背景复杂，仅靠肉眼分辨困难。目前虽然有多种方法可对野生菌进行分类，且最为可靠的方法为分子鉴定法，但该方法耗时长、门槛高，不适合进行实时分类检测。针对这一问题，提出了一种基于深度学习的方法，即使用注意力机制(CBAM)，配合多尺度特征融合，增加 Anchor层，利用超参数进化思想对其模型训练时的超参数进行调整，从而提升识别精度。与常见的目标检测网络 SSD，Faster_Rcnn 和 Yolo 系列等进行对比，该模型能更准确地对野生菌进行分类检测；经过模型改进后，相较于原Yolov5，Map 提升 3.7%，达到 93.2%，准确率提升 1.3%，召回率提升 1.0%，且模型检测速度提升 2.3%；相较于 SSD，Map 提升 14.3%。最终将模型简化，部署到安卓设备上，增加其实用性，解决当前因野生菌难以辨别而误食不可食用野生菌导致中毒的问题。

关键词: 计算机应用, 卷积神经网络, 多尺度特征, 超参数进化, 注意力机制, 可食用野生菌, 目标检测

Abstract:

In China, there are frequent poisoning events caused by ingestion of inedible wild mushrooms every summer, especially in Southwest China, such as Yunnan. This is due to the slight differences in inter-class characteristics of wild mushrooms and the complex image backgrounds in actual scenarios, making it difficult to distinguish only by naked eyes. At present, although there are many methods to classify wild mushrooms, and the most reliable way is molecular identification, the relevant techniques are time-consuming and require a high threshold, so they are not suitable for real-time classification and detection. To solve this problem, an approach based on deep learning was proposed. This approach employed the attention mechanism convolution block attention module (CBAM), was combined with multi-scale fusion, and added the anchor layer. The hyperparameter evolution idea was adopted to adjust the hyperparameter during the model training, so as to improve the recognition accuracy. Compared with standard target detection networks, such as SSD, Faster Rcnn, and Yolo series, the proposed model can classify and detect wild mushrooms more accurately. Compared with the original Yolov5, after the proposed model was improved, Map was improved by 3.7% and reached 93.2%, precision by 1.3%, Recall by 1.0%, and model detection speed by 2.3%. Compared with SSD, Map was improved by 14.3%. Finally, the model was simplified and deployed on Android devices to increase its practicability, thus solving the current problem of poisoning caused by eating inedible wild mushrooms because of the difficulty of identification.

Key words: computer application, convolutional neural network, multi scale features, hyperparameter evolution; attention mechanism, edible wild mushrooms, target detection 

中图分类号: