Research on Real-time Image Classification on Mobile Devices using Lightweight Convolutional Neural Networks Based on Deep Learning

Abstract

This dissertation is centered around the research on real-time image classification on mobile devices by means of lightweight convolutional neural networks grounded in deep learning. During the research process, initially, an in-depth analysis is carried out on a variety of extant lightweight convolutional neural network algorithms, taking into account their characteristics such as network architecture, parameter quantity, and computational complexity. Subsequently, in view of the characteristic of limited resources in mobile devices, methods such as optimizing the network architecture, reducing parameter redundancy, and adopting quantization techniques are employed to improve the selected lightweight convolutional neural network. In the image classification task, large-scale public image datasets are utilized for training and validation, and practical tests are conducted on mobile devices. The experimental results demonstrate that the improved lightweight convolutional neural network, while ensuring a high classification accuracy, significantly reduces the computational load and memory occupation, enabling real-time image classification on mobile devices. The processing speed meets the requirements of practical applications, providing a more efficient solution for image classification applications on mobile devices.