PERFORMANCE–EFFICIENCY TRADE-OFFS OF LIGHTWEIGHT DEEP LEARNING ARCHITECTURES FOR REAL-TIME COMPUTER VISION ON RESOURCE-CONSTRAINED DEVICES: A LITERATURE-SYNTHESIZED COMPARATIVE ANALYSIS

Abstract

The deployment of deep convolutional neural networks on mobile phones, embedded boards, and other resource-constrained platforms is increasingly limited not by classification accuracy alone but by the joint budget of inference latency, memory footprint, model size, and computational cost measured in floating-point operations (FLOPs). This paper investigates how five widely adopted lightweight architecture families — MobileNet (V1–V3), ShuffleNet (V1–V2), SqueezeNet, EfficientNet (including EfficientNetV2), and the YOLO-nano/-tiny family of real-time object detectors — negotiate the trade-off between predictive performance and computational efficiency for real-time computer vision. Rather than training new models on new hardware, this study adopts a literature-synthesis methodology: it systematically collects, cross-verifies, and normalizes accuracy, inference-time, model-size, parameter-count, and FLOPs figures that have already been measured and published by the original authors of each architecture and by independent comparative benchmarking studies, most notably a 2025 cross-dataset evaluation of MobileNetV3, ShuffleNetV2, SqueezeNet, EfficientNetV2, and ResNet18 on CIFAR-10, CIFAR-100, and Tiny ImageNet. The research is motivated by practical experience building real-time, camera-based computer-vision applications — including an AI-assisted exercise-monitoring system based on human pose estimation and a facial emotion-recognition prototype — in which inference speed on ordinary consumer hardware was as critical as raw accuracy. The synthesis identifies a consistent pattern: architectures that minimize parameter count and FLOPs (SqueezeNet, ShuffleNetV2) achieve the fastest inference and smallest footprint but suffer disproportionate accuracy degradation on complex, fine-grained recognition tasks, whereas EfficientNet-family models achieve the highest accuracy at the cost of larger memory footprints and slower inference, with MobileNetV3 consistently occupying the most balanced position across datasets of increasing difficulty. For real-time object detection, YOLO-nano and YOLO-tiny variants reproduce this same pattern at the detection level, trading mean average precision for frames-per-second suitable for edge boards such as the Raspberry Pi and Jetson Nano. The paper contributes a multi-dimensional, literature-grounded comparison framework, a research-gap analysis showing that most prior comparative studies restrict themselves to one or two efficiency metrics, and practical deployment guidance for engineers selecting architectures under specific latency, memory, or power constraints. Because no new experimental hardware measurements were collected by the authors, all quantitative claims are explicitly traced to their original published source, and the paper distinguishes clearly between findings drawn from the literature and the authors' own comparative synthesis.