基于深度神经网络的树木伐桩轮廓提取及匹配方法  

Research on contour extraction and matching method ofstump based on deep neural network

作　　者：崔世林[1] 田斐[1] CUI Shilin;TIAN Fei(Intelligent Manufacturing College,Nanyang Institute of Science and Technology,Nanyang 473004,China)

机构地区：[1]南阳理工学院智能制造学院,南阳473004

出　　处：《林业工程学报》2024年第3期149-158,共10页Journal of Forestry Engineering

基　　金：河南省重点科技攻关项目(172102210414)。

摘　　要：为了及时准确地找到被盗树木,公安机关需要比对被盗伐树木伐桩的上下截面,寻找共同点,并依此确认两者是否属于同一树木。但是由于存放环境不同,伐桩上下截面的颜色、纹理存在巨大差异,由于锯伐方式和树皮的影响,伐桩上下表面的轮廓也存在很大差异,伐桩下表面还很容易受木屑等影响,背景复杂。针对这些难点,本研究在前面的研究基础上,继续把棋盘格作为特征物放置在伐桩表面,用PPYOLO_MobileNetV3卷积神经网络检测图像中的棋盘格,对棋盘格中的角点进行检测、排序,然后进行透视变换,恢复伐桩的原始面积和轮廓等特征,接着用PP-LiteSeg网络在复杂背景下提取伐桩轮廓,然后用CAE_ViT_base网络对轮廓进行匹配,实现了伐桩轮廓匹配的全流程,从而极大程度节约了人工。理论分析和试验结果都表明,基于局部梯度的匹配法、基于局部点集拓扑特征的匹配法和基于轮廓的全局特征匹配方法等,在伐桩图像的匹配中都是不可行的。CAE_ViT_base网络的解码器将输入图像分割为大小一致的图像块,解码器的训练过程需要关注每一个块的特征,伐桩轮廓的匹配难点在于轮廓有局部缺失,局部梯度误差较大。CAE_ViT_base网络的自监督预训练机制很好地弥补了上述缺点;同时,采用对样本图像随机多角度旋转的方法,使得图像的特征提取能够保持旋转不变性。CAE_ViT_base网络提取出来的特征在尺度上高于基于梯度的特征,也高于基于局部点集拓扑的特征,但低于全局特征。因此,只要少部分图像块高度匹配,则CAE_ViT_base网络给出的最终匹配度就比较高;同时,这种工作方式和人工对比2个伐桩轮廓是否匹配的方法也是一致的。在本研究的344幅伐桩图像上进行试验,结果表明:本研究算法对整个测试集的检测成功率为100%,棋盘格检测成功率100%,轮廓提取精度达到98.8%,轮廓匹配准确率100%,�In order to find stolen trees timely and accurate,the public security organs need to compare the upper and lower surfaces of the stumps by searching for their common ground and confirming whether the two stumps belong to the same tree.However,due to different storage environments,there are significant differences in the color and texture of the upper and lower surfaces of the stumps.Due to the influence of felling method,there are also significant differences in the contour of the upper and lower of the stumps.The lower surface of the stump is also easily affected by wood chips and difficult to clean.In response to these difficulties,based on the previous research,this study used the chessboard as a feature and placed it on the surface of the stump,and used the PPYOLO_MobileNetV3 convolutional neural network to detect the chessboard in the images.After the corners in the chessboard were detected and sorted,a perspective transformation was performed to restore the original area and contour of the stump,and then used PP-LiteSeg network to extract the stump contour under the complex background,and then used CAE_ViT_base network to implement the stump contour matching.CAE_ViT_base network divided the input images into uniformly sized image blocks,and the training process of the decoder needed to pay attention to the characteristics of each block.The difficulty in matching the contour of the stump lies in the presence of local missing contours,and large local gradient errors.CAE_ViT_base network used the self supervised pre-training mechanism,which can effectively compensate for the above shortcomings.At the same time,a method of randomly rotating the sample images from multiple angles was adopted,so that the feature extraction of the images can maintain rotation invariance.The features extracted by this mechanism were higher in scale than gradient based features and local point set topology-based features,but lower than global features.The final matching degree given by the CAE_ViT_base network was relatively high,wh

关 键 词：深度神经网络 棋盘格检测 伐桩轮廓检测 轮廓匹配 伐桩校正 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP181[自动化与计算机技术—计算机科学与技术]