以卷積神經網路為基礎之改良型可解釋性深度學習模型;An Improved CNN-Based Interpretable Deep Learning Model

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/98493

請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98493

題名:	以卷積神經網路為基礎之改良型可解釋性深度學習模型;An Improved CNN-Based Interpretable Deep Learning Model
作者:	陳俊宇;Chen, Chun-Yu
貢獻者:	資訊工程學系
關鍵詞:	可解釋人工智慧;深度學習;色彩感知;彩色影像;Explainable Artificial Intelligence;Deep Learning;Color Perception;Color Images
日期:	2025-08-04
上傳時間:	2025-10-17 12:50:34 (UTC+8)
出版者:	國立中央大學
摘要:	隨著深度學習模型在醫學影像與電腦視覺等高風險應用領域取得優異表現，其「黑箱性」亦日益受到關注。為提升模型的透明度與可解釋性，本研究以 2024 年提出之 RGB CNN-based Interpretable Model（RGBCIM）為基礎，針對其關鍵模組進行多項改良。首先，在彩色卷積模組方面，我們棄用傳統 PCCS 色環，改採 CIELAB 色彩空間中，經 K-means 聚類後均勻分布之 30 種濾波器，以更貼近人眼感知的方式強化濾波器對色彩特徵之表徵能力。其次，在高斯卷積模組中，我們引入餘弦相似度作為卷積運算基礎，於保有相似度意涵並提升準確度的同時，大幅降低對超參數微調之敏感度。最後，在可解釋性流程整合方面，我們延續原 RGBCIM 之視覺化流程，提出濾波器監測指標，並新增 Grad-CAM 篩選機制，以產出更清晰且聚焦之解釋圖。本研究於 Colored MNIST、Colored Fashion MNIST、Colored Shape、PathMNIST、BloodMNIST、CIFAR-10、RetinalMNIST 等七個資料集上進行實驗驗證，新模型在所有資料集上的分類準確率皆優於原始 RGBCIM，平均提升幅度顯著。同時，面對複雜背景時亦能產出更具辨識力之可解釋性圖像，充分證實本研究改良方案在模型精度與可解釋性之間達成良好平衡。;With the outstanding performance of deep learning models in high-risk application domains such as medical imaging and computer vision, increasing attention has been drawn to their "black-box" nature. To enhance model transparency and interpretability, this study builds upon the RGB CNN-based Interpretable Model (RGBCIM) proposed in 2024 and introduces several improvements to its key modules. First, in the color convolution module, we replace the traditional PCCS color circle with 30 uniformly distributed filters obtained through K-means clustering in the CIELAB color space. This design better aligns with human color perception and enhances the filters’ ability to represent color features. Second, in the Gaussian convolution module, we adopt cosine similarity as the basis for convolution operations. This not only preserves the semantic meaning of similarity and improves accuracy but also significantly reduces the sensitivity to hyperparameter tuning. Lastly, in the integration of the interpretability pipeline, we extend the original RGBCIM’s visualization process by proposing a filter monitoring metric and introducing a Grad-CAM-based filtering mechanism to produce clearer and more focused explanation maps. Experiments conducted on seven datasets — Colored MNIST, Colored Fashion MNIST, Colored Shape, PathMNIST, BloodMNIST, CIFAR-10, and RetinalMNIST — demonstrate that the improved model achieves higher classification accuracy across all datasets compared to the original RGBCIM, with significant average gains. Additionally, it generates more distinguishable interpretability maps when faced with complex backgrounds, confirming that the proposed improvements successfully strike a balance between model accuracy and interpretability.
顯示於類別:	[資訊工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	18	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....