基於CNN之輕量化GAN圖像修復網路設計

Abstract

隨著深度學習模型日趨複雜，其高昂的計算需求限制了在資源受限設備上的應用。本論文旨在設計並驗證一高效輕量化的卷積神經網路(CNN)架構，研究中以圖像修復(Image Inpainting)作為具體應用場景，其核心目標在於顯著降低模型的參數數量與計算複雜度，同時維持優異的修復品質。 本研究以公開高品質人臉資料集 FFHQ 作為資料來源，並選定輕量化模型 MI-GAN 作為改進的基礎架構。研究方法的核心在於提出一種由雙參數 (α, β) 控制的卷積通道削減策略。此策略受 GhostNet 與 ShuffleNet 啟發，將恆等映射(identity mapping)整合至深度卷積與逐點卷積層中。此策略僅對一部分特徵通道執行卷積運算，其餘通道則直接保留。此設計不僅將計算資源集中於關鍵特徵之提取，更透過結構性約束降低了卷積核之間的餘弦相似度，從而隱性地促進權重正交性，提升了訓練穩定性。 實驗結果顯示，本研究所提出的改良模型 (MI-GAN-(Ours), α=0.85, β=0.1) 相較於原始模型，在生成器參數量與記憶體存取次數(MAC) 分別降低了11.67%與13.3%，推論延遲減少了54.54%。儘管達到相近修復品質 (以 FID 分數衡量) 所需的訓練迭代次數增加了 60%，但本研究證實了該方法能在推論效率、模型複雜度與生成品質之間取得有效的權衡，此成果為輕量化圖像修復網路的設計提供了一套具體且可行的架構優化方案。

As deep learning models become increasingly complex, their high computational demands limit their application on resource-constrained devices. This paper aims to design and validate an efficient and lightweight convolutional neural network (CNN) architecture for image inpainting tasks based on generative adversarial networks (GANs). The core objective is to significantly reduce the model's parameter count and computational complexity while maintaining excellent inpainting quality. This study utilizes the public high-quality face dataset, FFHQ, as the data source and selects the lightweight model MI-GAN as the foundational architecture for improvement. The core of the research methodology lies in proposing a convolutional channel reduction strategy controlled by dual parameters (α, β). Inspired by GhostNet and ShuffleNet, this strategy integrates identity mapping into depthwise and pointwise convolutional layers. The strategy performs convolution operations on only a portion of the feature channels, while the remaining channels are preserved directly. This design not only focuses computational resources on the extraction of critical features but also implicitly promotes weight orthogonality by reducing the cosine similarity between convolutional kernels through structural constraints, thereby enhancing training stability. The experimental results show that the proposed improved model (MI-GAN-(Ours), α=0.85, β=0.1), compared to the original model, reduces the number of generator parameters and memory access cost (MAC) by 11.67% and 13.3%, respectively, and decreases inference latency by 54.54%. Although the number of training iterations required to achieve a comparable inpainting quality (as measured by the FID score) increased by 60%, this study demonstrates that the proposed method achieves an effective trade-off between inference efficiency, model complexity, and generation quality. This result provides a concrete and viable architectural optimization scheme for the design of lightweight image inpainting networks.