Weakly supervised object localization (WSOL) is recognized as a promising yet challenging task that aims to achieve object localization using only image-level category labels. However, existing Class Activation Mapping (CAM) methods are limited to using low-resolution deep feature representations of the network, where only the most discriminative parts of the target are highlighted, resulting in incomplete localization and poor boundary precision. To overcome these limitations, a High-Resolution Class Activation Mapping (HR-CAM) algorithm is proposed based on the Contrastive Layer-wise Relevance Propagation (CLRP). The algorithm can be seamlessly integrated into convolutional neural networks, enabling high-quality class activation maps to be generated without modifying the backbone architecture. To fully exploit the low-, mid-, and high-level feature representations of the network, a hybrid additive–multiplicative fusion strategy is designed. Additive fusion is used to aggregate multi-level features to preserve fine-grained object details, while multiplicative refinement is applied to leverage high-level semantic activations to suppress background noise and enhance salient regions. Experimental results show that high-quality class activation maps are generated by HR-CAM, enabling accurate and efficient object localization and validating its effectiveness and superiority.