<![CDATA[Lightweight convolutional neural networks (CNNs) are increasingly required for image classification in resource-constrained environments; however, their comparative behavior under unified training conditions remains insufficiently explored, particularly when accuracy, parameter efficiency, inference latency, and augmentation sensitivity are evaluated simultaneously. This study presents a systematic benchmark of five lightweight CNN architectures, namely MobileNetV2, EfficientNet-B0, ShuffleNetV2, SqueezeNet, and ResNet18, on the CIFAR-100 dataset using a consistent experimental pipeline. All models were trained for 40 epochs with an input resolution of 128 × 128, AdamW optimization, cosine annealing, mixed-precision training, and identical preprocessing settings. Two augmentation strategies, namely basic and advanced augmentation, were evaluated to examine their influence on model generalization. The results show that EfficientNet-B0 achieved the best classification performance with 82.75% Top-1 accuracy and 96.46% Top-5 accuracy, while SqueezeNet achieved the fastest inference latency of 1.52 ms and the smallest parameter size, indicating its suitability for highly constrained deployment scenarios. Across all evaluated models, the average Top-1 and Top-5 accuracies reached 76.6% and 94.16%, respectively. In addition, the effect of advanced augmentation was found to be architecture-dependent rather than uniformly beneficial. On average, it resulted in a Top-1 accuracy change of −0.66 percentage points, with only ResNet18 showing a modest improvement. The main contribution of this study is to provide a unified, practically oriented benchmark that highlights how architectural design, rather than parameter count alone, determines the balance between accuracy and computational efficiency. These findings provide clearer guidance for selecting lightweight CNN models for real-world image classification tasks under varying deployment constraints.]]>
