Deep Neural Networks for Vision Tasks on Mobile Devices (Seminar)
Abstract
Deep neural networks (DNNs) have propelled significant advancements in computer vision tasks, resulting in excellent and noticeable performance in the fields of image classification, object detection, semantic segmentation, and image generation. With the widespread use of mobile devices, there is a growing demand to deploy these DNNs directly on mobile platforms. However, deploying DNNs on resource-constrained devices such as smartphones gives rise to unique challenges because of limited computational resources, power constraints, network bandwidth, and memory limitations. This seminar report describes optimization techniques and advancements in adapting deep neural networks for mobile devices. The focus is on understanding the adaptations and optimization techniques employed to balance computation and performance without compromising accuracy. The report surveys various neural network backbones, including MobileNet, ShuffleNet, SqueezeNet, EfficientNet, MobileOne, and Mobile-Former, highlighting their architectural features and efficiency strategies. The practical implementation of these architectures in various applications substantiates their efficacy in Computer Vision tasks. This report provides valuable insights for selecting and implementing vision-based models on consumer devices by exploring the optimization approaches and advancements in resource-optimized deep neural networks.
Topic
A survey of vision-based deep neural network architectures designed for resource-constrained consumer devices such as smartphones. In this seminar, you will explore different neural network backbones and how they optimize computation and performance on these devices without losing accuracy.
Tasks
- Learn about various backbones proposed for deep learning on mobile phones.
- Explore how these neural networks are optimized to work in resource-constrained environments.