Golden Phoebe wood texture imagery data and its training process with convolutional neural networks

The samples A to E were processed uniformly into rectangular prisms with dimensions of 80 mm × 200 mm × 15 mm. To expand the number of samples, the upper rectangular prism samples were sectioned, resulting in 95 images per sample. Ultimately, we obtained a total of 475 images (comprising the A-E 5 major categories) with an average imagery of different wood textures (W-A1-W-E95). These were used as the foundation for constructing a CNN-based imagery recognition model for Phoebe wood, which combines users' affective semantic evaluations of Phoebe wood (such as "Nature-Artificial," "Variety-Single") to link visual image characteristics with users' subjective perceptions. To meet the requirements of the input layer of the ResNet-152 model, each sample was divided into 15 images under the same lighting angle, resulting in a total of 475 × 15 = 7125 images of Phoebe wood grain patterns. To delve deeper into the subjective evaluation words and scores of Phoebe wood grain, and to extract the imagery features of Golden Phoebe wood, the entire dataset was randomly divided into a training set (96%) and a test set (4%), containing 6840 and 285 samples, respectively. Each training sample image was subjected to random horizontal flipping, vertical flipping, and rotation operations before being input into the convolutional neural network.The images used as training data were processed by randomly initializing the weights of the neural network, employing the MSE Loss function as the loss function, and undergoing linear regression training. The entire convolutional neural network model was trained for 35 epochs, and finally, the validation data was used for testing. Figure 10 depicts the graph of Loss rate changes during the training period. After 35 epochs of training, the convolutional neural network has reached a state of convergence, demonstrating the usability of the model.