In order to address the issue of the AlexNet network model's insufficient recognition accuracy and robustness in Chinese herbal slice image recognition, a study was conducted to investigate the improvement and optimization of the AlexNet network model. Firstly, images of 50 common Chinese herbal slices were obtained through shooting and search engines, and the images underwent data expansion and detail enhancement preprocessing. Then, the AlexNet network model was optimized and improved by lowering the original network's convolutional kernel number and size. Global average pooling (GAP) was used instead of full connection layer to reduce network parameters. The local response normalization (LRN) layer was removed and the batch normalization (BN) layer was introduced. The Lion optimizer was used to replace the stochastic gradient descent (SGD) optimizer to improve network training speed. The Mish activation function was used in place of ReLU activation function, and the channel attention mechanism SENet network was introduced to enhance the recognition accuracy of the model. The experimental results show that compared to the AlexNet network model, the improved network model exhibits an average recognition rate increase of 6.1%, an average loss rate decrease of 14.4% and the network parameters have been reduced from the original 60 M to 1 M. It is concluded that the improved network model shows higher recognition accuracy and better robustness on the dataset of Chinese herbal slices, providing strong support for further development in the field of image recognition of Chinese herbal slices.