python实现的基于多视图卷积神经网络的三维物体识别算法实现三维物体的准确识别源代码+文档说明.zip

# Multi-view-Convolutional-Neural-Networks-for-3D-Shape-Recognition 本文提出了一种基于多视图卷积神经网络的三维物体识别算法，以实现三维物体的准确识别。首先实现一个标准的卷积神经网络架构，该架构经过训练可以独立地识别形状的渲染视图，以实现即使从单一视图中也可以识别出一个三维形状。随后使用该三维物体多个角度的二维视图通过卷积神经网络识别的结果进行模型融合。在模型融合的过程中取出输入单角度视图的卷积神经网络的某一层，使用层最大值算法将多个层中同一位置的最大值取出，形成一个新的层参与神经网络的训练，以提高卷积神经网络识别多视图的准确率。为了应对数据集较小的情况，我们使用了迁移学习的方法，先使用ModelNET数据集对单视图卷积神经网络进行预训练，然后将预训练模型加载到多视图卷积神经网络后再进行微调，以提高图像的识别率。ModelNET数据集上的实验结果证明了算法的有效性。Keras版本：2.2.4，TensorFlow版本：1.12.0，GPU型号：GeForce RTX 2080 Ti In this paper, we propose a multi-view convolutional neural network which based on 3D object recognition algorithm to achieve accurate recognition of 3D objects. A standard convolutional neural network architecture is first implemented that is trained to recognize rendered views of a shape independently to achieve recognition of a 3D shape even from a single view. The 2D views of the 3D object from multiple angles are then used to perform ensemble learning with the results recognized by the convolutional neural network. A layer of the convolutional neural network with an input single-angle view is taked during the ensemble learning process, and the maximum value at the same location in multiple layers is taked using the layer maximum algorithm to form a new layer to participate in the training of the neural network to improve the accuracy of the convolutional neural network in recognizing multiple views. To cope with the small dataset, we use a transfer learning approach to pre-train the single-view convolutional neural network using the ModelNET dataset first, and then fine-tune the pre-trained model after loading it into the multi-view convolutional neural network to improve the recognition rate of images. experimental results on the ModelNET dataset demonstrate the effectiveness of the algorithm.Keras version: 2.2.4, TensorFlow version: 1.12.0, GPU model: GeForce RTX 2080 Ti