Enhanced MobileNets via Augmentation of Filtering-based Features

Thanks to the significant performance while maintaining a small model size, a MobileNet-based model is one of the potential solutions for mobile applications in image classification. However, taking into account raw image data can lead to a lack of structural information of the input images for the learning process of MobileNets. To this end, we propose to take advantage of structural patterns extracted by a well-known image filter. Firstly, for an input image, a popular filter will be addressed to capture its geomorphological features. Then, a simple function is executed to concatenate the filtered result with the initial image to form an augmented tensor. Thereby, instead of addressing the raw images, their corresponding augmented tensors will be fed into the learning process of MobileNets to exploit richer features for image description. Experimental results on benchmark datasets have validated the effectiveness of our proposal. Particularly, the performance of MobileNetV3 has been boosted by 2.5% on CIFAR-100 and up to 5% on Stanford Dogs, while the computational complexity seems to be the same. The implementation code of our proposal is available at https://github.com/quoctc/EnhancedMNet.