Few-Shot Infrared Image Classification with Partial Concept Feature

Few infrared image samples will bring a catastrophic blow to the recognition performance of the model. Existing few-shot learning methods most utilize the global features of object to classify infrared image. However, their inability to sufficiently extract the most representative feature for classification results in a degradation of recognition performance. To tackle the aforementioned shortcomings, we propose a few-shot infrared image classification method based on the partial conceptual features of the object. It enables the flexible selection of local features from targets. With the integration of these partial features into the concept feature space, the method utilizes Euclidean distance for similarity measurement to accomplish infrared target classification. The experimental results demonstrate that our proposed method outperforms previous approaches on a new infrared few-shot recognition dataset. It effectively mitigates the adverse effects caused by background blurring in infrared images and significantly improving classification accuracy.