Lightweight PhiNet-Based Neural Network for Image Classification on Edge Devices

Abstract

Many AI applications still require substantial computation and memory, making edge computing unsuitable for resource-limited devices. Therefore, this paper will design a relatively lightweight image classification framework based on the enhanced PhiNet architecture. Dynamically select filters and perform hierarchical quantization to achieve the best balance between accuracy and efficiency. In order to evaluate the framework, the benchmark datasets ImageNet-1K and CIFAR-100 were selected using the aforementioned standard experimental procedures. The above results indicate that the model based on PhiNet achieved a Top-1 accuracy of 77.0% on ImageNet-1K and 85.1% on CIFAR-100, surpassing the corresponding Top-1 accuracies of MobileNetV3 and EfficientNet-Lite. On the NVIDIA Jetson Xavier NX platform, the average inference latency per image is 23.1 milliseconds, with a throughput of 42.4 frames per second. In addition, the average power consumption is 2.7 watts. When batch processing up to 32 photos, the memory usage remains below 95 MB. In addition, many tests in various real-world environments have shown good consistency and robustness. These environments include recognizing pedestrians in low light and recognizing license plates in motion. Increase hardware-aware optimizations so that the system can more easily adapt to various hardware. Finally, this paper demonstrates that the PhiNet-based solution achieves an ideal balance in terms of accuracy, efficiency, and robustness, making it suitable for practical applications in edge intelligence.