研究目的
To identify and count the texture features in ancient buildings using smartphones for real-time detection, overcoming the challenges of complexity and large number of textures.
研究成果
The proposed ancient architecture texture and component detection technology applied to smart phones based on deep learning shows better effect on the features with larger dataset. Future studies can use data augment to increase the accuracy and achieve more functions.
研究不足
The recognition of Chandu is the worst due to its oblique part and small dataset. The method is not accurate for small training sets and has better effect on features with larger datasets.
1：Experimental Design and Method Selection:
Uses SSD-Mobilenet, a Convolutional Neural Network (CNN), for training.
2：Sample Selection and Data Sources:
180 samples of ancient building features, including Juancaowen, Dizuo, Yingzui, and Chandu.
3：List of Experimental Equipment and Materials:
Smartphones, TensorFlow framework, object detection API, SSD-Mobilenet v2 network.
4：Experimental Procedures and Operational Workflow:
Training the model using TensorFlow and object detection API, then importing the trained model into a smartphone for real-time detection.
5：Data Analysis Methods:
Evaluation and validation using 20 pictures not in the training set, with results viewed in Tensorboard.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
