Bronze inscriptions are rarely different from one another, deep learning network training is susceptible to overfitting, and the deeper the convolutional layer, the more intricate elements of the inscription are lost, leading to a low level of classification accuracy. A small sample bronze inscription classification approach that incorporates morphological aspects is offered in light of this issue. To preprocess the hole filling of inscriptions and lessen the impact of noise on the morphological structure of inscriptions, morphological algorithms must first be introduced. Second, modify the structure of the AlexNet network, and in each convolution layer to introduce batch normalization, control each batch of random input values, so that it conforms to the normal distribution standards, and avoid overlearning the network in a single direction to suppress overfitting. Last but not least, the speeded-up robust features (SURF) operator is used to extract the finer morphological details of the inscriptions. This merging of finer morphological details with abstract features from convolutional neural networks improves classifier expression. The classification accuracy in the bronze inscription dataset experiment is 98.86%, which is higher than traditional algorithms such as LeNet5, Vgg13, Vgg16, ResNet, and AlexNet and effectively addresses the issue of low classification accuracy.