| DC 欄位 |
值 |
語言 |
| DC.contributor | 資訊工程學系 | zh_TW |
| DC.creator | 童子祐 | zh_TW |
| DC.creator | Tzu-Yu Tung | en_US |
| dc.date.accessioned | 2022-7-30T07:39:07Z | |
| dc.date.available | 2022-7-30T07:39:07Z | |
| dc.date.issued | 2022 | |
| dc.identifier.uri | http://ir.lib.ncu.edu.tw:444/thesis/view_etd.asp?URN=109522114 | |
| dc.contributor.department | 資訊工程學系 | zh_TW |
| DC.description | 國立中央大學 | zh_TW |
| DC.description | National Central University | en_US |
| dc.description.abstract | 由於卷積神經網路模型擁有龐大的演算法，運算過程被視為黑盒子(black box)無法對其提出合理的解釋與分析，因此本研究提出透過增強影像特徵的方式並結合MLP-Mixer 分類器，增加整個辨識系統的可解釋性與準確，該辨識系統架構應用於魚類、種子和中歐森林生物辨識資料集。首先針對影像先進行形狀、紋理與顏色的特徵增強，再將特徵增強過後的影像(Feature-Enhanced Image, FEI)作為 MLPMixer 分類器的輸入，分別輸出三個特徵增強方式的 Top-5，作為此三個 Top-5 作為決策融合的輸入，透過多類別羅吉斯回歸(Multinomial Logistic Regression)輸出最終決策結果。本篇研究在 40 種魚類資料集上達到 99%的辨識率，優於未使用特徵增強的 MLP-Mixer 分類器的 96%辨識率；在 560 類種子資料集上達到 90.65%的辨識率，優於混合式神經網路(ResNet-50+Siamese)的 70.23%辨識率；在中歐森林資料集153 類上達到 97.91%的辨識率，優於採用單個卷積神經網路架構的 93.4%辨識率。 | zh_TW |
| dc.description.abstract | Since the convolutional neural network model has a huge algorithm, the operation process is regarded as a black box and cannot provide a reasonable explanation and analysis. Therefore, this study proposes to enhance the image features and combine the MLP-Mixer classifier to increase the overall Interpretability and accuracy of the identification system architecture applied to the fish, seed and central European forest biometric datasets.Firstly, the features of shape, texture and color are enhanced for the image, and then the image after feature enhancement is used as the input of the MLPMixer classifier, and the Top-5 of the three feature enhancement methods are output respectively, as the three Top-5 as the input of the MLP-Mixer classifier. The input of the decision fusion, the final decision result is output through multi-class Logis regression.This study achieves a recognition rate of 99% on 40 fish datasets, which is better than the 96% recognition rate of the MLP-Mixer classifier without feature enhancement;Achieving a recognition rate of 90.65% on the 560-category seed dataset, which is better than the 70.23% recognition rate of the hybrid neural network;It achieves a recognition rate of 97.91% on 153 categories of the Central European Forest dataset, which is better than the 93.4% recognition rate using a single convolutional neural network architecture. | en_US |
| DC.subject | 影像特徵增強 | zh_TW |
| DC.subject | 方向梯度直方圖 | zh_TW |
| DC.subject | 局部二值模式 | zh_TW |
| DC.subject | 單尺度視網膜增強算法 | zh_TW |
| DC.subject | image feature enhancement | en_US |
| DC.subject | Histogram of Oritentd Gradients | en_US |
| DC.subject | Local Binary Patterns | en_US |
| DC.subject | Single-Scale Retinex | en_US |
| DC.subject | MLP-Mixer | en_US |
| DC.title | 使用特徵增強策略在 MLP-Mixer 影像分類器 | zh_TW |
| dc.language.iso | zh-TW | zh-TW |
| DC.title | Applying Feature Enhancement Strategies in MLP-Mixer Image Classifier | en_US |
| DC.type | 博碩士論文 | zh_TW |
| DC.type | thesis | en_US |
| DC.publisher | National Central University | en_US |