This paper studies the classification problem of the small sample multi-label image scene recognition. Combining convolutional neural network (CNN) and multi-label K neighborhood algorithm (MLKNN), the CNN-MLKNN classification method is proposed. The method uses CNN to automatically extract the features of small sample images, and combines transfer learning to optimize the model structure and weight to reduce the risk of over-fitting. MLKNN algorithm is used to replace the sigmoid function of CNN, and the output features of the FC layer are used as input features of MLKNN for image classifier training. Based on the classification experiments of two small sample multi-label image sets, seven multi-label evaluation indicators are used for testing. The experimental results show that the CNN-MLKNN method proposed in this paper has a better classification effect.
展开▼