Integrated Mamba Block with Spatially-Aware Feature Fusion

Visual modeling in CNNs has been widely enhanced by various attention mechanisms designed to integrate channelwise and spatial-wise information, achieving prominent results across a range of vision tasks. However, despite their efficiency in enhancing local feature interactions, these designs remain rooted in convolutional operations, which constrain their capacity for global spatial reasoning. Inspired by Mamba, an efficient longrange selective state space model, we propose a lightweight Integrated Mamba Block (IMB) comprising a bidirectional Mamba branch and a depth-wise convolution branch. These two pathways are combined through a spatially-aware fusion mechanism that exchanges channel descriptors and generates fine-grained spatial attention, which enables effective integration of global and local features with an acceptable computational cost. We conduct extensive experiments on CIFAR-100 with ResNet-50, demonstrating consistent performance gains and supporting the effectiveness of each component through ablation studies. The module also demonstrates transferability to lightweight architectures such as MobileNetV2, highlighting its general applicability. Code is available at https://github.com/undetectedatom/mamba-fusion-module