A Review of Online Learning Methods for Weighted Fusion of Visible Light-Infrared Images for Edge Intelligence

Addressing the core requirement of real-time adaptive adjustment of fusion weights in visible–infrared cross-modal fusion tasks for edge intelligence scenarios, this paper systematically reviews the research progress in this field from four dimensions: cross-modal feature fusion, online convex optimization, lightweight neural networks, and neural architecture search. First, the dynamic fusion weight allocation methods based on attention mechanisms are summarized, and the applications of cross-attention, channel attention, and self-attention mechanisms in mining modal complementarity are analyzed; comparative experiments are conducted to reveal the performance differences among various fusion strategies in terms of key metrics such as information entropy, standard deviation, and spatial frequency. Second, the theoretical framework of online convex optimization and its mathematical foundation for real-time weight updating are elaborated, and the regret bound analysis under gradient-free optimization and bandit feedback scenarios is discussed, establishing a bridge between theoretical guarantees and engineering practice. Third, lightweight neural network architecture design and model compression techniques are summarized, including knowledge distillation, network pruning, and quantization methods; the computational efficiency and accuracy trade-offs of mainstream architectures such as SqueezeNet, MobileNet series, and EfficientNet are comparatively analyzed. Finally, hardware-aware neural architecture search for automated network design under ultra-low computational power constraints is reviewed. On this basis, three core challenges facing current research are thoroughly analyzed: heterogeneous modal feature alignment and semantic consistency modeling, the coupled optimization of “hardware–model–real-time–accuracy”, and the online learning mechanism under ultra-low computational power constraints; future research directions such as neuro-symbolic fusion, continual learning, and hardware–software co-design are also prospected.