Comparison of chosen image classification methods on Android

The authors compared three lightweight convolutional networks (MobileNet-V1, EfficientNet-Lite0 and ResNet-50) for image classification on Android smartphones using TensorFlow Lite and a multithreaded CPU. They measured inference time, CPU load and memory usage across various devices. EfficientNet-Lite0 proved the best compromise - providing high accuracy, short and consistent processing times and moderate resource demands. MobileNet-V1 was the fastest but less precise, while ResNet-50 achieved the highest accuracy at the expense of speed and memory. In practice, EfficientNet-Lite0 is recommended, and further research into optimizations such as quantization, pruning and adaptive frame sampling is advised.