Yanju Meng, Peng Wu, Jian Feng, XiaoMing Zhang
Currently, vision transformers (ViTs) have rivaled comparable performance to convolutional neural networks (CNNs). However, the computational demands of the transformers’ self-attention mechanism pose challenges for their application on edge devices. Therefore, in this study, we propose a lightweight transformer-based network model called MixMobileNet. Similar to the ResNet block, this model only comprises a MixMobile block (MMb), which combines the efficient local inductive bias with the explicit modeling features of a transformer to achieve the fusion of the local–global feature interactions. For local, we propose the local-feature aggregation encoder (LFAE), which incorporates a PC2P (Partial-Conv→PWconv→PWconv) inverted bottleneck structure for residual connectivity. In particular, the kernel and channel scale are adaptive, reducing feature redundancy in adjacent layers and efficiently representing parameters. For global, we propose the global-feature aggregation encoder (GFAE), which employs a pooling strategy and computes the covariance matrix between channels instead of the spatial dimensions, changing the computational complexity from quadratic to linear, and this accelerates the inference of the model. We perform extensive image classification, object detection, and segmentation experiments to validate model performance. Our MixMobileNet-XXS/XS/S achieves 70.6%/75.1%/78.8% top-1 accuracy with 1.5 M/3.2 M/7.3 M parameters and 0.2 G/0.5 G/1.2 G FLOPs on ImageNet-1K, outperforming MobileViT-XXS/XS/S with an improvement of +1.6%↑/+0.4%↑/+0.4%↑ with −38.8%↓/−51.5%↓/−39.8%↓ reduction in FLOPs. In addition, the MixMobileNet-S assembly of SSDLite and DeepLabv3 achieves an accuracy of 28.5 mAP/79.5 mIoU at COCO2017/VOC2012 with lower computation, demonstrating the competitive performance of our lightweight model.