一种基于改进SSD的原木端面识别方法  

Development of log end face recognition method based on improved SSD

作　　者：胡笑天 王克俭[1,2,3] 王超 剪文灏[4] 何振学 HU Xiaotian;WANG Kejian;WANG Chao;JIAN Wenhao;HE Zhenxue(College of Information Science and Technology,Agriculture University of Hebei,Baoding 071001,China;Hebei Urban Forest Health Technology Innovation Center,Baoding 071001,China;Key Laboratory of Agricultural Big Data of Hebei Province,Baoding 071001,China;State-owned Forest Farm Administration Bureau of Mulanweichang,Hebei Province,Chengde 067000,China)

机构地区：[1]河北农业大学信息科学与技术学院,保定071001 [2]河北省城市森林健康技术创新中心,保定071001 [3]河北省农业大数据重点实验室,保定071001 [4]河北省木兰围场国有林场管理局,承德067000

出　　处：《林业工程学报》2023年第1期141-149,共9页Journal of Forestry Engineering

基　　金：河北省自然科学基金(F2020204003);河北省高等学校科学技术研究项目(BJ2019008,ZD2016158)。

摘　　要：自然环境下的原木楞堆摆放参差不齐,端面大小不一,且存在遮挡、重叠、被阴影覆盖等现象,识别效果不理想。为了提高原木端面识别率,笔者对原木端面识别方法进行了研究。采用深度学习的方法,针对混楞堆原木端面大小差距较大、较小原木难以检测以及遮挡情况下不容易提取有效特征等问题,以速度较快的SSD(single shot multibox detector)网络为基础网络,对conv_fc7使用上采样后与conv4_3融合替换原conv4_3,将conv4_3和conv_fc7两个有效特征层进行结合,增加了感受野,提高了该特征层对较小原木端面的特征提取能力;在整体结构上加入融合多尺度卷积核和空洞卷积的RFB(receptive field block)模块,又在网络中引入能够使网络学习特征权重加强对有效特征关注的CBAM (convolutional block attention module),增强了特征识别能力。实验使用从原木验收现场采集的图像,结果表明:改进SSD目标检测网络对清楞原木的检测精确率达96.37%,召回率96.81%,AP值99.06%;在含有较多小目标原木的混楞测试集检测中改进SSD检测精确率97.00%,召回率92.90%,AP达到95.33%,召回率比SSD提高了14.03%。改进SSD网络增强了SSD目标检测网络的抗干扰能力,扩大了感受野,提高了原木端面实时检测性能。Due to the uneven log piles in the natural environment, the sizes of the log end faces are very different, and some log end faces are blocked, overlapped, or covered by shadows, which leads to unsatisfactory recognition results. In order to improve the recognition rate of log end faces, this study analyzes the log end faces by a cutting-edge technology. The identification method was developed using the improved single shot multibox detector(SSD) method. Based on the deep learning technology, in order to solve the problems such as the large difference in the size of the end faces of the mixed logs, the difficulty of detecting small logs, and the difficulty in extracting effective features under occlusion, a faster SSD network is used as the basic network. After up sampling conv_fc7, it is merged with conv4_3 to replace the original conv4_3, and the two effective feature layers of conv4_3 and conv_fc7 are combined to increase the receptive field and improve the feature extraction ability of this feature layer on the end faces of small logs. The receptive field block(RFB) module that integrates multi-scale convolution kernels and cavity convolutions is added to the overall structure, and the convolutional block attention module(CBAM) that enables the network to learn feature weights and strengthen attention to effective features is introduced into the network. Module is used to enhance the feature recognition capability. The experiment is based on the images collected from the log acceptance site. The results show that the improved SSD target detection network has a detection accuracy of 96.37%, a recall rate of 96.81%, and an average precision(AP) value of 99.06%. In the integrated testing, the accuracy rate of SSD detection is 97.00%, the recall rate is 92.90%, and the AP reaches 95.33%, the recall rate is 14.03% higher than that of SSD. The improved SSD network enhances the anti-interference ability of the SSD target detection network, expands the log pile receptive field, and improves the real-time detection per

关 键 词：原木识别 SSD目标检测网络 RFB模块 注意力模块 

分 类 号：S126[农业科学—农业基础科学]