Spiking neural networks for object detection and semantic segmentation across event-driven and frame-based modalities: a review

Spiking neural networks (SNNs), drawing inspiration from the energy-efficient and event-driven processing of biological brains, are emerging as a compelling alternative to traditional artificial neural networks (ANNs) for resource-constrained artificial intelligence (AI) applications. Their intrinsic properties, including low power consumption, ultra-low latency, and native spatio-temporal information processing capabilities, position them as ideal candidates for critical computer vision tasks such as real-time object detection and semantic segmentation, especially at the edge. This review systematically explores the fundamental principles of SNNs, including their unique neuron models and information encoding schemes, contrasting them with the operational paradigms of ANNs. We delve into the sophisticated mathematical formulations underpinning key SNN neuron models and the intricate learning dynamics that differentiate SNNs. A significant portion is dedicated to meticulously dissecting recent architectural innovations in SNNs tailored for image object detection and semantic segmentation. This includes an in-depth analysis of pure SNN convolutional networks, pragmatic hybrid SNN-ANN models, and the cutting-edge integration of attention mechanisms and Transformer-based designs. Furthermore, we provide an enhanced exposition of crucial training algorithms, such as advanced surrogate gradient methods and spiking batch normalization, highlighting their theoretical underpinnings and practical implications. Finally, this review synthesizes the current performance benchmarks, identifies persistent research challenges, and delineates promising future directions, particularly emphasizing the synergistic co-design of SNN algorithms and neuromorphic hardware. We argue that SNNs, while not yet universally outperforming ANNs, hold immense potential to revolutionize AI in dynamic, resource-limited environments, becoming a cornerstone of next-generation intelligent systems.