Abstract: Aimed at improving the performance of the depth model in image classification currently, i.e. the inadequate performance of existing hardware, difficulty in structural innovation and the limited training samples, an object fusion recognition algorithm based on DSmT(Desert-Smarandache theory)was proposed. The recognition information of objects was collected and fused from different learning network models. The pretrained depth learning models were fine-tuned according to the classification task. To solve the problem in the construction of the basic belief assignment(BBA)in DSmT, the models were used to assign the BBA to the evidence sources. The DSmT combination theory was used in the fusion of the decision-layer in order to raise the recognition rate. Under the conditions of unchanged network models and the dataset, the multi-model fusion method with the single-model and average value method were compared in the experiments. The results of the experiments showed that the algorithm could improve correct recognition ratio effectively under the same conditions.

Key words: deep learning, information fusion, object recognition, Dezert-Smarandache theory, deep neural network