Convolutional Neural Network Image Classification Method Integrating Classic Image Features

Classic image features were once widely used in image classification but have been almost entirely replaced by neural networks today. While the performance of neural networks, especially convolutional neural networks (CNNs), is indisputable, their lack of interpretability has become a significant limitation in recent years. This paper explores the effective combination of classic image features and convolutional image features, investigating their differences and similarities. Through a series of experiments, a method is proposed to effectively integrate these two types of features by employing a multiplicative attention mechanism. This approach combines standardized MPEG-7 descriptors with convolutional features before feeding them into the fully connected layer for classification. The final model demonstrates an improvement in classification accuracy and indicates the potential of learning traditional image features from the integrated features.