Spiking Neural Networks (SNNs) have attracted significant attention due to their biological plausibility, exceptional spatiotemporal processing capabilities, and low power consumption. Nevertheless, despite these advantages, SNNs face challenges in efficiently handling dynamic image datasets, particularly due to issues related to computational inefficiencies and the complex nature of temporal data. Existing methods encounter difficulties in effectively integrating multi-scale features and optimising spiking neurons for environments with limited resources. The present work proposes the Reversed Fusion Spiking Neural Network (RFSNN), an innovative architecture designed to address these limitations. The incorporation of reversed fusion and spatial-temporal attention mechanisms within the RFSNN framework enhances the fusion of temporal and spatial features across multiple scales, thereby improving the accuracy of dynamic image recognition while concurrently reducing computational expenditure. This approach represents a substantial advancement in the efficiency and scalability of SNNs, providing a robust solution for real-time applications in dynamic environments. The effectiveness of RFSNN is validated through extensive experimentation on dynamic datasets, including DVS-Gesture, CIFAR-10, and MNIST, demonstrating its capacity to achieve high recognition performance with minimal resources. This advancement paves the way for more efficient neuromorphic computing solutions.