This paper studies the application of spiking neural networks (SNNs), characterized by low power consumption and fast inference on neuromorphic hardware, to visual recognition tasks. Existing ANN-to-SNN conversion methods have shown excellent performance in classification tasks, but underperform in visual detection tasks. To address this, this paper proposes a delayed-spike approach and a time-dependent integrate-fire (tdIF) neuron architecture that mitigates residual membrane potential issues caused by heterogeneous spiking patterns. tdIF neurons dynamically adjust their accumulation and firing behaviors according to the order of time steps, enabling spikes to exhibit distinct temporal characteristics without relying on frequency-based representations. Furthermore, they maintain energy consumption comparable to that of conventional IF neurons. Extensive evaluations on two visual tasks—object detection and lane detection—show that the proposed method outperforms existing ANN-to-SNN conversion methods, achieving state-of-the-art performance in fewer than five time steps.
