Computer Engineering and Applications ›› 2018, Vol. 54 ›› Issue (23): 89-93.DOI: 10.3778/j.issn.1002-8331.1806-0224

Previous Articles     Next Articles

Identification method of ambrostoma quadriimpressum motschlsky based on Faster R-CNN

DONG Benzhi, NIE Lili, JING Weipeng, CUI Hang   

  1. College of Information and Computer Engineering, Northeast Forestry University, Harbin 150040, China
  • Online:2018-12-01 Published:2018-11-30

基于Faster R-CNN的榆紫叶甲虫识别方法研究

董本志,聂丽郦,景维鹏,崔  航   

  1. 东北林业大学 信息与计算机工程学院,哈尔滨 150040

Abstract: Aiming at the problem that the traditional image recognition method uses artificially designed feature extraction templates to reduce the recognition accuracy of insects, the deep learning network model Faster R-CNN based on k-means clustering is proposed to identify the targets in the image. Firstly, it uses the k-means clustering algorithm with the BWP index to cluster the aspect ratios of the training data labels, and replaces the aspect ratios of the initial candidate box generated by the standard Faster R-CNN network with the new clustering center point. Then, it improves the size of the initial candidate box. Finally, it sends the training data to the improved Faster R-CNN network for training. The experimental result shows that the Faster R-CNN network with clustering strategy has stronger robustness than the standard Faster R-CNN network when identifying targets with a specific aspect ratio, effectively overcoming the redundancy caused by blade break or holes, the interference of the crustacean of the ambrostoma quadriimpressum motschlsky(aqm), the mutual influence of the adjacent aqm features and other types of insects with similar characteristics of the aqm. The final recognition accuracy is up to 94.73%. The final recognition accuracy of 94.73% is 4.15% higher than that of the standard network. The method can effectively overcome the limitations of the feature extraction template in traditional insect detection, and has important significance for the identification of insects with such features as delicate and varied attitude.

Key words: ambrostoma quadriimpressum motschlsky, insect recognition, convolutional neural network, Faster R-CNN, initial candidate box adjustment, k-means clustering algorithm

摘要: 针对传统图像识别方法中利用人工设计特征提取模板对昆虫的识别精度不高的问题,提出了基于K-means聚类的深度学习网络模型Faster R-CNN对图像中的目标进行识别。该方法用K-means聚类算法,结合BWP指标对训练数据标签的长宽比值进行聚类,用新的聚类中心点代替标准Faster R-CNN网络中生成初始候选框的长宽比值;对生成初始候选框的尺寸加以改进;将训练数据送入改进后的Faster R-CNN网络进行训练。实验结果表明,在识别具有特定长宽比例的目标时,加入聚类策略的Faster R-CNN网络较标准Faster R-CNN网络有较强的鲁棒性,有效克服了叶片豁口或孔洞造成的冗余现象、榆紫叶甲虫甲壳反光的干扰、相邻的榆紫叶甲虫特征的互相影响和其他与榆紫叶甲虫有相似特征的种类昆虫的干扰。最终达到94.73%的识别精度,较标准网络提高了4.15%。该方法可有效克服传统昆虫检测中特征提取模板的局限性,对识别昆虫这种特征细腻,姿态多样的目标有重要意义。

关键词: 榆紫叶甲虫, 昆虫识别, 卷积神经网络, Faster R-CNN, 初始候选框调整, K-means聚类算法