Year: 2023
Type: article
Abstract: On-device computer vision now depends on convolutional architectures deliberately engineered to fit within the memory, latency, and power budgets of phones, embedded boards, and microcontrollers, rather than repurposed from server-scale networks. This article synthesizes the literature that produced those architectures: depthwise separable convolutions, popularized by Xception and operationalized for mobile deployment by the MobileNet family; the channel-shuffle and inverted-residual-bottleneck refinements introduced by ShuffleNet and MobileNetV2; the platform-aware neural architecture search that produced MnasNet, MobileNetV3, and the EfficientNet compound-scaling family; and the complementary, largely orthogonal literature on pruning, quantization, and knowledge distillation that compresses these architectures further once trained. We give particular attention to the most recent phase of this work, from 2021 onward, in which structural reparameterization (RepVGG, MobileOne) and hybrid convolution-attention designs (MobileViT, EdgeNeXt, FastViT) have extended the same efficiency-first design philosophy to compete with, and selectively borrow from, vision transformers. Read together, the literature supports lightweight convolutional design as a cumulative, empirically validated research program rather than a series of disconnected tricks, while leaving genuinely open the questions of how architecture, quantization, and target hardware should be jointly optimized, and whether pure convolutional, pure transformer, or hybrid designs will dominate the next generation of on-device vision models.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7213645805