In complex military visual recognition scenarios, targets are easily affected by camouflage, occlusion, cluttered backgrounds and low-quality imaging. Meanwhile, large intra-class pose variations and high inter-class structural similarity often lead to prototype deviation under few-shot conditions, while first-order distance metrics are insufficient for modeling fine-grained distribution differences. To address the problem that prototype construction and similarity measurement are usually processed in a serial manner without effective feedback collaboration, a Prototype and Metric Collaboration Network, termed PMCN, is proposed. In PMCN, frequency-domain attention and channel discriminative enhancement are designed to suppress background pseudo-textures and strengthen category-related channels. A Differentiable Iterative Prototype Optimizer, named DIPO, is employed to refine category prototypes according to sample contribution and feature consistency. Furthermore, a covariance-based metric learning module, named CML, is constructed to model second-order correlation structures between query samples and category prototypes. Bidirectional collaborative optimization is achieved by using CML feedback to correct prototypes and DIPO reliability to modulate metric representation. Experimental results show that PMCN outperforms mainstream few-shot learning methods on ME-35, BS-30 and mini-ImageNet. In particular, accuracies of 68.9% and 80.1% are achieved on ME-35 under 5-way 1-shot and 5-way 5-shot settings, respectively, demonstrating the effectiveness of the proposed method.