合成孔徑雷達（Synthetic Aperture Radar， SAR）影像因具備全天候與全天時成像能力，已成為海事監控與船舶偵測中不可或缺的資料來源。然而，SAR 影像在成像過程中會受到相干成像特性影響而產生斑點雜訊（speckle noise），該雜訊以乘法形式存在，容易降低影像對比度、模糊目標邊界，並干擾後續自動化影像分析與目標偵測作業。近年來，深度學習技術在電腦視覺領域快速發展，其中以 Faster R-CNN 為代表的目標偵測模型已成功應用於 SAR 船舶偵測任務，但其偵測效能仍高度仰賴輸入影像品質，斑點雜訊的存在對模型訓練與推論結果仍構成顯著挑戰。 本研究針對 SAR 船舶偵測流程中斑點雜訊抑制之影響進行系統性與定量化分析，旨在比較不同斑點雜訊濾波方法作為前處理步驟時，對深度學習偵測效能所造成的差異。研究選用四種常見之自適應斑點雜訊濾波器，包括 Lee 、Frost 、Kuan 及 GammaMap 濾波器，並以未進行任何濾波處理之 SAR 影像作為基準組。為確保實驗比較之公平性與可重現性，所有實驗皆採用相同之資料集、訓練與驗證資料切分方式，以及一致之模型架構與訓練參數設定。 本研究以 ResNet-50 作為特徵擷取骨幹，建構 Faster R-CNN 船舶偵測模型，並分別於不同斑點雜訊處理條件下進行模型訓練與效能評估。模型效能分析涵蓋訓練過程之收斂行為，以及驗證階段之平均準確率（Mean Average Precision， mAP）等指標，以全面評估各濾波方法對偵測表現之影響。實驗結果顯示，斑點雜訊抑制方法的選擇對深度學習模型之訓練穩定性與最終偵測效能具有顯著影響。相較於未濾波影像，適當的斑點雜訊抑制可有效改善模型收斂表現並提升整體偵測準確率；然而，過度平滑可能削弱船舶之邊界與結構特徵，反而不利於偵測任務。整體而言，GammaMap 濾波器在雜訊抑制與目標特徵保留之間展現較佳平衡，於多數評估指標中取得相對優異之表現。本研究透過實證分析說明斑點雜訊抑制在 SAR 船舶深度學習偵測流程中的重要性，研究成果可作為未來相關研究與實務應用在前處理策略選擇上的參考依據。;Synthetic Aperture Radar (SAR) imagery supports maritime surveillance with all-weather， day-and-night coverage. Yet coherent imaging introduces speckle noise that lowers contrast and blurs ship boundaries， challenging deep learning detectors. This study quantitatively evaluates how speckle-noise reduction affects SAR ship detection using Faster R-CNN. Four adaptive speckle filters—Lee， Frost， Kuan， and GammaMap—are evaluated as preprocessing strategies， and unfiltered imagery serves as the baseline. For a controlled comparison， all experiments use the same dataset， identical training/test split， and fixed model architecture and training hyperparameters. A Faster R-CNN detector with a ResNet-50 backbone is trained for each preprocessing condition and tested on the same test set. Performance is assessed by training convergence behavior and test detection accuracy， using mean Average Precision (mAP) as the primary metric. Results show that preprocessing choice significantly influences both training stability and final detection performance. Moderate speckle suppression generally improves convergence and increases mAP， while excessive smoothing can remove salient ship structures and weaken boundary cues， reducing detection effectiveness. Among the evaluated methods， GammaMap provides the most favorable balance between noise reduction and feature preservation and achieves the best overall performance across most metrics. These findings highlight preprocessing as a critical design choice in SAR ship detection pipelines and provide empirical guidance for selecting speckle-noise reduction methods in deep learning–based maritime monitoring applications. In particular， filters that reduce speckle while preserving scatterers and sharp edges tend to support more consistent region proposals and classification scores.
