MobileGhost: Lightweight Ghost Module for Image Classification

The ever-growing ubiquity of mobile devices has fueled the demand for image classification models that are both accurate and computationally efficient. While deep convolutional neural networks (CNNs) excel in accuracy, their high resource demands often hinder their deployment on resource-constrained devices. This work introduces MobileGhost, a novel lightweight image classification architecture that achieves remarkable robustness by combining the strengths of GhostNet and MobileOne. MobileGhost leverages GhostNet’s “cheap operation” principle, replacing standard convolutions with linear bottlenecks that significantly reduce computational cost while maintaining representational power. This is further enhanced by incorporating residual skip connections and squeeze-and-excitation modules, resulting in a model with exceptional feature reuse and channelwise attention. These combined innovations enable MobileGhost to achieve superior accuracy compared to other lightweight models, even on challenging datasets with complex visual patterns. Our evaluations demonstrate that MobileGhost consistently outperforms baseline models in terms of accuracy and robustness against various corruptions and transformations. Furthermore, MobileGhost boasts exceptional speed and memory efficiency, making it ideally suited for real-world deployment on mobile devices.