DeepMine-Mamba: Mitigating Information Dilution in Mamba-Based State Space Models for Document Image Binarization

Document binarization has to separate text from backgrounds carrying uneven illumination, bleed-through, stains, faded ink, and paper texture. Mamba's linear-time propagation is attractive here, because high-resolution manuscripts need both global degradation context and local stroke precision.

But document images are mostly background. When state-space propagation runs across a page, background-dominant responses overwhelm weak foreground cues, and faint ink traces, fragmented characters, and boundary-sensitive stroke details get diluted along the way. This failure mode is specific to the data — it is not a generic accuracy problem — and it had not been named before.

文件二值化必須把文字從帶有不均勻光照、透印、污漬、褪色墨跡與紙張紋理的背景中分離出來。 Mamba 的線性時間傳播在這裡很有吸引力，因為高解析度的手稿同時需要全域的劣化脈絡與局部的筆畫精度。

但文件影像的絕大部分是背景。當狀態空間傳播掃過整頁時，背景主導的響應會淹沒微弱的前景線索， 淡墨痕跡、斷裂的字元、對邊界敏感的筆畫細節，就在傳播途中被稀釋掉。 這個失效模式是資料特性造成的——不是一般性的準確度問題——而且先前沒有被命名過。

The key move is to measure the damage rather than guess at it. Given the original local feature x and the Mamba-propagated feature m, the term |x − m| describes exactly how much each location changed during propagation. The Anti-Dilution Gate feeds that difference into a lightweight detector to produce a dilution importance map, then modulates it by a stroke prior so compensation is confined to stroke regions instead of amplifying background noise. The output interpolates between the two: y = m + Î ⊙ (x − m). When the gate is closed, the model keeps Mamba's global context; when it opens, local stroke response comes back.

Around this sit a ConvNeXt-Tiny encoder, a Sobel edge prior fused late for detail-aware prediction, four-directional Mamba propagation with directional fusion, and three prediction heads that estimate foreground, background, and an adaptive threshold map. Training uses a pseudo F-measure loss computed on the skeletonised ground truth, which penalises missing stroke cores directly — the same objective the anti-dilution design is chasing.

關鍵作法是去量測損失，而不是猜測損失。給定原始的局部特徵 x 與經 Mamba 傳播後的特徵 m，|x − m| 這一項正好描述了每個位置在傳播過程中被改變了多少。 Anti-Dilution Gate 把這個差值送入輕量偵測器，產生稀釋重要性圖，再用筆畫先驗加以調製， 使補償只作用在筆畫區域，而不會連背景雜訊一起放大。輸出則在兩者之間插值： y = m + Î ⊙ (x − m)。閘門關閉時，模型保留 Mamba 的全域上下文； 閘門開啟時，局部筆畫響應就被找回來。

圍繞這個核心的是：ConvNeXt-Tiny 編碼器、在後段融合以支援細節感知預測的 Sobel 邊緣先驗、 帶方向性融合的四方向 Mamba 傳播，以及分別估計前景、背景與自適應閾值圖的三個預測頭。 訓練使用在骨架化標註上計算的 pseudo F-measure 損失，直接懲罰遺漏的筆畫核心—— 這正是 anti-dilution 設計所追求的目標。