Abstract: Aiming at the problem that single-modal object recognition is easily affected by climate and interference, an object recognition algorithm based on multi-modal image feature fusion is proposed. The proposed method is divided into the following four primary steps. In the first step, multi-modal images are preprocessed, during which the visible light, infrared, and SAR images of the ground scene are augmented respectively based on their modality-specific characteristics to enhance the model′s generalization ability and robustness. Next, corresponding classification labels are assigned to each image. Subsequently, different deep learning networks are designed to train data with different modalities separately, and the classification heads of all pre-trained networks are then removed. Finally, each classification network is connected to the feature fusion module for a retraining, so as to improve the classification accuracy of the object classification. The method proposed in this paper accomplishes information complementation by employing multi-modal image feature fusion for object recognition, thereby achieving efficient identification of space objects. Experimental results demonstrate that the proposed algorithm achieves a high recognition accuracy of 96.74 % on the WHU-OPT-SAR dataset.