Application of lightweight convolutional neural networks in image classification

Abstract

In recent years, the demand for real-time and resource efficiency in image classification tasks has been growing, which has promoted the rapid development of lightweight convolutional neural networks (CNNs). This paper selects four representative CNN models - ResNet-18, MobileNetV3, ShuffleNetV2 and EfficientNet-B0, and systematically evaluates their performance and deployment adaptability in image classification from the perspectives of standard training and small sample training based on the CIFAR-10 dataset. The experiment compares them from multiple dimensions such as accuracy, macro-average F1 score, convergence speed, number of parameters, computational complexity (FLOPs) and inference latency. The results show that ResNet-18 performs best under the full dataset conditions, but ShuffleNetV2 shows the most stable and balanced performance under small sample scenarios. EfficientNet-B0 shows significant generalization improvement after introducing data augmentation (especially AutoAugment), and obtains good classification results while maintaining moderate computational overhead; while MobileNetV3, although it has faster convergence and lower resource usage, is limited by its representation ability and has weaker performance under small sample conditions. This study further investigated model selection recommendations for different deployment scenarios, such as embedded systems, robotic platforms, and server-side inference. The results showed the importance of co-designing model structure and data augmentation strategy in alleviating overfitting and improving model generalization capabilities.