THFOOD-50 (Thai Food 50 Image Classification)

目前，卷积神经网络（CNN）已被广泛应用于许多应用中。图像识别是利用 CNN 的应用之一。对于该领域的大部分研究，CNN 主要用于提高识别的有效性。但是，处理时间和参数量（或模型大小）并未作为主要因素考虑。本文研究了使用智能手机对泰国食品进行图像识别。减少了处理时间和模型大小，以便它们可以与智能手机正确使用。论文提出了一种新的网络，称为 NUInNet（Naresuan University Inception Network），它采用了 GoogLeNet 中使用的 Inception 模块的概念。使用名为THFOOD-50的泰国食品数据库进行应用和测试，该数据库包含50种著名的泰国食品。结果发现，与从 GoogLeNet 获得的相比，NU-InNet 可以将处理时间和模型大小分别减少 2 倍和 10 倍，同时保持识别精度与 GoogLeNet 相同。使用所提出的网络显着减少处理时间和模型大小肯定可以满足智能手机中泰国食品识别应用的用户。

Currently, Convolutional Neural Networks (CNNs) have been widely adopted across a diverse range of applications. Image recognition is one of the prominent use cases for CNNs. For most research in this field, CNNs are primarily optimized to improve recognition performance, while processing time and model parameters (or model size) are seldom taken as core design considerations. This paper focuses on Thai food image recognition using smartphones, aiming to reduce both processing time and model size to enable proper deployment on mobile devices. We propose a novel convolutional neural network named NU-InNet (Naresuan University Inception Network), which adopts the Inception module concept from GoogLeNet. Experiments and evaluations are conducted using the THFOOD-50 Thai food database, which contains 50 well-known Thai food categories. The results demonstrate that, compared to the baseline GoogLeNet model, NU-InNet can reduce processing time and model size by 2-fold and 10-fold respectively, while maintaining identical recognition accuracy to GoogLeNet. The significant reductions in processing time and model size achieved by the proposed network can undoubtedly meet the user requirements of Thai food recognition applications on smartphones.

搜集汇总

背景与挑战

背景概述

THFOOD-50是一个用于泰国食品图像分类的数据集，包含50种著名泰国食品的图像。该数据集旨在支持智能手机应用中的食品识别研究，通过NU-InNet网络优化处理时间和模型大小，同时保持与GoogLeNet相当的识别精度。