<p>In order to reduce the dependence of the model training process on labeled sample data and further develop low-cost, high-performance traditional recognition methods, a semi-supervised decision-level fusion bag-of-word (SD-BoW) is proposed by combining three improvements based on BoW and extreme learning machine (ELM). To balance training time and performance, the dictionary vector generation process is improved. A small number of dictionary vectors are generated through clustering based on the K-means method, and most dictionary vectors are generated through a combination of random selection and average calculation. Aiming at the problem of enhancing the diversity information of representation-level features at multiple resolutions, a multi-resolution feature encoding channel is constructed. Unlabeled data is utilized based on the guidance of labeled category information. Inter-class information is calculated separately in different resolution channels for autoencoding training. Encoding coefficients of different resolutions are combined with pooled features to generate diverse representation-level features. On the basis of the above improvements, a semi-supervised decision-level fusion ELM (SDELM) is constructed to address the problem of lack of available information in the semi-supervised training process. By fusing semi-supervised and decision-level constraints to improve the objective function of ELM, a new semi-supervised method is solved to train classification weights. To verify the feasibility of SD-BoW, experiments are conducted on the Olympic Sports, UCF11, UCF101, MNIST, and NORB databases. Experimental results show that the proposed SD-BoW, while achieving semi-supervised training, can further integrate and utilize the identification information between features of different resolutions to obtain excellent recognition accuracy compared with other methods.</p>