Improving ShuffleNet and SqueezeNet Using Residual Network for Small Device

The development of technology, particularly hardware, has been very rapid lately. Small devices, which previously required limited memory and processor speed, now have capacities and performance comparable to mid-range devices. As a result, some deep learning research that was too focused on minimal size needs to be revised. Some deep learning models suitable for small capacities but needing performance enhancement have become the focus of this research. Two models, SqueezeNet and ShuffleNet, will enhance their performance by inserting residual networks (ResNet) to improve performance. To objectively determine the effects of adding the ResNet module, both the data and hyperparameters, e.g., learning rate, number of epochs, etc., are kept the same. Data is taken from benchmark data with 978 training images across 9 classes. The test results show that the combination of ShuffleNet with ResNet outperforms SqueezeNet with ResNet.