The traditional metric learning based few-shot image classification methods are task independent, which leads to poor generalization performance of the model on new query tasks. To solve this problem, a taskrelevant image few-shot learning method was proposed in this paper, which can adaptively adjust the feature of support samples according to the query task. Moreover, a variety of regularization methods to address the overfitting problem under severely-limited data scenarios were also investigated. We conduct comprehensive experiments on two popular benchmarks, i.e., miniImageNet and tieredImageNet. The result of 1-shot task on the miniImageNet by the proposed method was 66.05%, and it outperforms the SOTA (state of the art) approaches by 4.29% under the same backbones.