基于深度学习的 5 种树皮纹理图像识别研究

本文以 5 种常见树种的树皮纹理图像为例，采用基于卷积神经网络的深度学习方法，将原 始图像直接作为输入，通过卷积和池化层对图像的低级、高级特征进行自动提取，解决了手动提取纹理特征的困难和问 题；在此基础上，对 CNN 模型结构进行改进，采用带 Maxout 的 ELU 激励函数来代替 ReLU 函数，解决模型的偏移和零 梯度问题；对损失函数进行改进，通过添加规范项来优化结构参数，并使用分段常数衰减法对学习率进行动态调控；最后 采用 softmax 分类器对图像类别进行输出。

This study takes bark texture images of 5 common tree species as examples, adopts a deep learning method based on convolutional neural networks (CNNs), directly uses the original images as input, and automatically extracts low-level and high-level features of the images through convolutional and pooling layers, solving the difficulties of manually extracting texture features. On this basis, the structure of the CNN model is improved: the ELU activation function with Maxout is employed to replace the ReLU function, which addresses the issues of model bias and vanishing gradient. Furthermore, the loss function is optimized by adding regularization terms to refine the structural parameters, and the piecewise constant decay method is utilized to dynamically adjust the learning rate. Finally, the softmax classifier is applied to output the category of each image.

搜集汇总

背景与挑战

背景概述

该数据集采用基于卷积神经网络的深度学习方法，对5种常见树种的树皮纹理图像进行智能识别，解决了手动提取纹理特征的困难。数据集属于'东北天然次生林抚育更新技术研究与示范'项目，数据格式为文档，数据量为363.27 KB，共享级别为协议共享数据。