Abstract: 【Purpose/significance】Aiming at the problem of insufficient expression of image semantic features in the current computer
management of image resources, the impact of features and feature fusion on the classification results is explored and analyzed, and a method to improve the accuracy of image semantic classification is proposed.【Method/process】This paper defines four image styles, divides image description features into three levels, explores the characteristics of feature fusion, and seeks features that can effective? ly express image semantics. SVM, CNN, LSTM and transfer learning methods are used to realize image style classification, and the al? gorithms are combined to improve the classification performance.【Result/conclusion】The deep feature extracted by the ResNet18 model based on transfer learning can better express the high-level semantics of the image, and combining it with SVM can improve the classification accuracy. Features are not always complementary, we should avoid feature redundancy when selecting features, which will reduce the efficiency of classification.【Innovation/limitation】The number of styles defined in the article is not enough, and the styles displayed by the image is not absolute, which can often be given multiple labels. In the future, the image data set should be fur? ther enriched and multi-label classification experiments should be tried.