Research on Bone Stick Text Recognition Method with Multi-Scale Feature Fusion

Bone sticks are composed of thin slices of animal bones created by ancient people, which mainly served the functions of fixing books, writing scripts, and divination. The bone stick script is an essential material for studying the history of Chinese Western Han script. Using a neural network for text recognition can quickly interpret ancient text, while extracting deeper semantic information, neural networks will also lose superficial image details. After multi-layer convolution and pooling of bone sticks, the continuous loss of superficial details affects classification accuracy. At the same time, the unbalanced distribution of bone stick quantity leads to a low recognition rate with small samples of bone sticks. Aiming to solve the above problems, a bone stick recognition method based on multi-scale features and focal loss function is proposed. Firstly, based on the residual network ResNet, the output features of the first layer and four Conv_x layers are pooled globally to reduce the feature dimension of each channel, and the channel splicing method is used to add different depths of base information to the original high-level features, which improves the detail feature extraction ability of the model. Secondly, in view of the unbalanced distribution of the bone stick data, the original cross-entropy loss function is replaced by the focus loss function, which increases the penalty for classification errors and improves the recognition rate of classes with few training samples. Experimental results show that the recognition accuracy of the proposed method on the bone stick data set is up to 90.5%.