The goal of on-device continual learning is to enable models to adapt to streaming data without forgetting previously acquired knowledge, even with limited computational resources and memory constraints. Recent research has demonstrated that weighted regularization-based methods are constrained by indirect knowledge preservation and sensitive hyperparameter settings, and dynamic architecture methods are ill-suited for on-device environments due to increased resource consumption as the structure scales. In order to compensate for these limitations, replay-based continuous learning, which maintains a compact structure and stable performance, is gaining attention. The limitations of replay-based continuous learning are (1) the limited amount of historical training data that can be stored due to limited memory capacity, and (2) the computational resources of on-device systems are significantly lower than those of servers or cloud infrastructures. Consequently, designing strategies that balance the preservation of past knowledge with rapid and cost-effective updates of model parameters has become a critical consideration in on-device continual learning. This paper presents an empirical survey of replay-based continual learning studies, considering the nearest class mean classifier with replay-based sparse weight updates as a representative method for validating the feasibility of diverse edge devices. Our empirical comparison of standard benchmarks, including CIFAR-10, CIFAR-100, and TinyImageNet, deployed on devices such as Jetson Nano and Raspberry Pi, showed that the proposed representative method achieved reasonable accuracy under limited buffer sizes compared with existing replay-based techniques. A significant reduction in training time and resource consumption was observed, thereby supporting the feasibility of replay-based on-device continual learning in practice.

Crack segmentation plays a vital role in ensuring structural safety, yet practical deployment on resource-limited platforms demands models that balance accuracy with efficiency. While high-accuracy models often rely on computationally heavy designs to expand their receptive fields, recent lightweight approaches typically delay this expansion to the deepest, low-resolution layers to maintain efficiency. This design choice leaves long-range context underutilized, where fine-grained evidence is most intact. In this paper, we propose an evidence-preserving receptive-field expansion network, which integrates a multi-scale dilated block to efficiently capture long-range context from the earliest stages and an input-guided gate that leverages grayscale conversion, average pooling, and gradient extraction to highlight crack evidence directly from raw inputs. Experiments on six benchmark datasets demonstrate that the proposed network achieves consistently higher accuracy under lightweight constraints. Each of the three proposed variants—Base, Small, and Tiny—outperforms its corresponding baselines with larger parameter counts, surpassing a total of 13 models. For example, the Base variant reduces parameters by 66% compared to the second-best CrackFormer II and floating-point operations by 53% on the Ceramic dataset, while still delivering superior accuracy. Pareto analyses further confirm that the proposed model establishes a superior accuracy–efficiency trade-off across parameters and floating-point operations.