人工智能
计算机科学
可解释性
增采样
模式识别(心理学)
计算机视觉
卷积神经网络
深度学习
特征提取
多层感知器
特征(语言学)
排列(音乐)
频道(广播)
面子(社会学概念)
图像处理
任务(项目管理)
代表(政治)
面部识别系统
图像(数学)
空间分析
小波
上下文图像分类
噪音(视频)
预处理器
可视化
人工神经网络
移动设备
Glyph(数据可视化)
失真(音乐)
作者
Mengjian Zhang,Tao Yang,Changjun WANG,Chuyun Chen,Kongfa Hu,Guihua Wen
标识
DOI:10.1007/s44443-026-01414-x
摘要
According to traditional Chinese medicine (TCM) theory, a well-defined correspondence exists between the five zang and six fu organs and specific manifestations observed in facial and tongue images. By integrating artificial intelligence technologies to investigate disease location using Inspection images, it becomes possible not only to enable precise auxiliary diagnosis but also to provide an auxiliary research reference for intelligent TCM multi-label disease location recognition based on retrospective inspection images of facial and tongue images. Disease location recognition (DLR) can be regarded as a multi-label classification task using Inspection images from TCM and computer vision theories. We propose a Permutation Wavelet-Inspired Multilayer Perceptron (PWI-MLP) architecture tailored for multi-label disease location recognition (DLR) using Inspection images. A convolutional tokenizer (Conv-Tokenizer) is employed to extract high-quality shallow features from input images, effectively compensating for the spatial information loss typically associated with a conventional Multilayer Perceptron (MLP). Subsequently, a channel permutation module and a channel interaction module are jointly designed to further capture deep contextual representations. Between successive stages, a discrete wavelet downsampling (DWDS) module is introduced, which enhances the frequency-domain representation of the extracted deep features while simultaneously reducing the overall number of model parameters. Finally, the resulting multi-channel fused features are utilized to perform multi-label DLR on the input Inspection images. Experimental results on the newly constructed Facial Image Multi-label Disease Location (FIMDL) and Tongue Image Multi-label Disease Location (TIMDL) datasets demonstrate that the proposed PWI-MLP model achieves significant performance advantages and provides visual interpretability consistent with TCM inspection theory in the multi-label DLR task. Besides, the performance of the PWI-MLP is also verified on a multi-label Ocular diseases dataset.
科研通智能强力驱动
Strongly Powered by AbleSci AI