MobileACNet: ACNet-Based Lightweight Model for Image Classification.

Lightweight CNN models aim to extend the application of deep learning from conventional image classification to mobile edge device-based image classification. However, the accuracy of lightweight CNN models currently is not as comparable as traditional large CNN models. To improve the accuracy of mobile platform-based image classification, we propose MobileACNet, a novel ACNet-based lightweight model based on MobileNetV3 (a popular lightweight CNN for image classification on mobile platforms). Our model adopts a similar idea to ACNet: consider global inference and local inference adaptively to improve the classification accuracy. We improve the MobileNetV3 by replacing the inverted residual block with our proposed adaptive inverted residual module (AIR). Experimental results show that our proposed MobileACNet can effectively improve the image classification accuracy by providing additional adaptive global inference on three public datasets, i.e., Cifar-100 dataset, Tiny ImageNet dataset, and a large-scale dataset ImageNet, for mobile-platform-based image classification.