BOCCHI 資料集與 MSDCT-UNet：局部動態模糊偵測

Abstract

局部動態模糊偵測旨在以像素級精度定位影像中受運動模糊影響的區域。當模糊與清晰區域可由梯度幅值簡單區分時，現有的模糊偵測基準容易讓模型依賴低梯度的代理特徵，而非真正學到運動模糊背後的物理性質。 為填補此空白，本研究建立 BOCCHI（Blurred Objects Captured across Cameras with Human-annotated Imagery）資料集：共 633 張以五台分屬不同廠牌的消費級相機在真實場景拍攝、並完成像素級人工標註的影像。BOCCHI 的清晰區域呈現雙峰梯度分布，其中低梯度模態（清晰 PR25 = 20.15）落在模糊區平均梯度（µblur = 29.62）之下，使清晰 PR25/µblur 比值（0.68）為受評五個資料集中最高，梯度分布重疊使單純依賴梯度幅值的判別捷徑失效。 我們進一步提出 MSDCT-UNet（Multi-Scale Discrete Cosine Transform UNet），一個對影像梯度進行多尺度 DCT 變換以取得頻域先驗、以多頭 DCT Attention 學習各頻率成分的重要性、並以 FiLM 機制融合頻域與空間特徵的編碼器-解碼器網路。在 BOCCHI 上，MSDCT-UNet 較第二名 baseline 提升 1.25 個百分點的 mIoU 並在邊界偵測上領先。跨資料集評估進一步揭示一個資料集層級的訓練效率優勢：以 BOCCHI 訓練的模型在 mIoU / Dice / Recall 平均值上勝過其他三個訓練來源（ReLoBlur、CUHKmotion、OMoBlur），且僅使用 633 張訓練影像，遠少於 ReLoBlur（1,200）與 OMoBlur（994）。此優勢與模型架構無關，顯示主導跨資料集可遷移性的關鍵是場景組成，而非訓練資料規模。

Local motion blur detection requires pixel-level localization of blurred regions. Existing benchmarks can permit gradient-based shortcuts when blurred and sharp regions are easily separable, letting models rely on low-gradient proxies rather than the underlying physics of motion blur. To address this gap, we introduce BOCCHI (Blurred Objects Captured across Cameras with Human-annotated Imagery), 633 real-captured, pixel-annotated images shot with five consumer-grade cameras spanning five different manufacturers. BOCCHI’s sharp regions exhibit a bimodal gradient distribution including a low-gradient mode (sharp PR25 = 20.15) that falls below the blur-region mean (µblur = 29.62), giving the largest sharp PR25/µblur ratio (0.68) among five evaluated datasets and creating gradient overlap that defeats simple gradient shortcuts. We further propose MSDCT-UNet (Multi-Scale Discrete Cosine Transform UNet), a frequency-aware encoder-decoder that extracts multi-scale DCT priors from gradients, weights them via multi-head DCT Attention, and fuses frequency and spatial cues through FiLM modulation. On BOCCHI, MSDCT-UNet surpasses the second-best baseline by 1.25 percentage points in mIoU and leads in boundary localization. Cross-dataset evaluation further reveals a dataset-level data-efficiency advantage: BOCCHI-trained models outperform every other training source (ReLoBlur, CUHKmotion, OMoBlur) on mIoU/Dice/Recall averages using only 633 training images, substantially fewer than ReLoBlur (1,200) and OMoBlur (994). This advantage is independent of architecture, suggesting that scene composition, rather than dataset scale, drives cross-dataset transferability for local motion blur detection.