A Deep Neural Network model for Sentiment Classification in Heilongjiang Tourism Images

The rapid development of Heilongjiang tourism has put forward an urgent demand for tourists' sentiment analysis. In this paper, we propose an optimization framework based on ResNet50, combining random oversampling, data enhancement and weighted loss strategies, to address the problems of uneven distribution of emotion categories and insufficient generalisation of complex scenes in tourism images. By comparing the performance difference between ResNet50, EfficientNet-B4 and MobileNetV2, the significant advantage of ResNet50 in classification accuracy is verified with a test accuracy of 99.71%. The ablation experiments further showed that random oversampling is a core component in solving category imbalance, and the model accuracy decreased to 77.28% after removal. This study provides an efficient engineering solution for tourism image sentiment analysis and quantifies the contribution of each technical component, which has important practical application value.