Detection

This guide covers converting an input image into text regions, a detection mask, and detection results for OCR. OCR, text-line merging, mask refinement, and inpainting belong to their respective settings pages. This guide focuses on the controls in the Detection tab, detection-stage behavior, and the annotation/debug files it actually uses.

Change it in the desktop app

In the desktop application, open “Settings” and select the “Detection” tab (the tab title is hard-coded in the layout and has no locale replacement). Basic items appear first; Advanced separates the size, threshold, and YOLO parameters. The dynamic settings page creates a combo box, toggle switch, or numeric input according to the current configuration type. Finishing an edit or changing a switch updates the in-memory configuration and schedules a merged config-file write.

Numeric inputs are submitted when they lose focus. Invalid numeric input does not become a valid core configuration value. Row labels are mapped by app_logic.py through the label_* keys above; configuration keys and environment-variable names must not be presented as UI labels.

Parameters

For the mapping of UI names, storage keys, and default values of the parameters on this page, see the Settings Parameter Index.

Text Detector

The “Text Detector” combo box is on Settings → Detection and selects the main detector for text regions.

• default: default detector.
• dbconvnext: DBNet-family detector.
• ctd: manga-oriented detector.
• craft: general-purpose text detector (not recommended for manga).
• none: runs no detector and returns no detections.

Default: default.

Detection Size

“Detection Size” is an integer input that sets the size the detector uses to scale the image. Larger values generally retain more small-text detail at the cost of compute and memory; an excessive value can exhaust VRAM. Default: 2048.

Long Image Rearrange Min Short Side

“Long Image Rearrange Min Short Side” is an integer input. When a long image is rearranged or tiled for detection, this value constrains the minimum effective short-side resolution; raising it keeps text clearer but costs more compute and memory. Default: 341.

Text Threshold

“Text Threshold” is a floating-point input. A higher text-confidence threshold is stricter and may miss text; lowering it retains weaker candidates and may add false positives. It controls the text response separately from “Box Generation Threshold”. Default: 0.5.

Box Generation Threshold

“Box Generation Threshold” is a floating-point input that sets the confidence gate for turning response candidates into text boxes. Lower values generally retain more boxes; higher values are stricter. Too low a value sends noise to OCR. Default: 0.5.

Unclip Ratio

“Unclip Ratio” is a floating-point input that controls how far a box expands from the text skeleton. Larger values generally make boxes larger and may include more background or neighboring text. Default: 2.5.

Import Fixed YOLO Boxes

When the “Import Fixed YOLO Boxes” toggle is enabled, same-name annotations are read from manga_translator_work/yolo_labels/. The template-plus-save-text flow can bypass the main detector and use imported boxes; the normal flow replaces detected boxes with imported ones. Label files must match the image name and the expected format. Default: false.

Enable YOLO Detection

When the “Enable YOLO Detection” toggle is enabled, the main detector runs first and is followed by an oriented YOLO detector; containment and overlap decide whether boxes are replaced, removed, or appended, and the main result is the fallback when auxiliary detection fails. It requires the YOLO model; the import-translation flow forces it off. Default: true.

YOLO Confidence Threshold

“YOLO Confidence Threshold” is a floating-point input used by YOLO assistance as its text-confidence gate. Raising it removes weak YOLO boxes; lowering it adds candidates and merging work. It does not replace the main detector’s “Text Threshold”. Default: 0.4.

YOLO Overlap Removal Threshold

“YOLO Overlap Removal Threshold” is a floating-point input that decides overlap handling when YOLO boxes merge with main-detector boxes. At or above the threshold, containment and area decide replacement or removal; below it, a YOLO box may be appended. Default: 0.1.

SFX Filter

“SFX Filter” does not filter text inside bubbles: bubble text is protected by the bubble mask by default and is only filtered when “Include Bubble Text in SFX Filter” is enabled. When enabled, main-detector boxes that are neither fully wrapped by a YOLO other box nor overlapped by a non-other YOLO box at the configured threshold are filtered; it depends on “Enable YOLO Detection”. Default: false.

Include Bubble Text in SFX Filter

When the “Include Bubble Text in SFX Filter” toggle is off, text inside a bubble that YOLO does not support remains protected by the bubble mask; enabling it skips that exemption and bubble text is filtered too. Enabling it can remove bubble dialogue. Default: false.

Min Box Area Ratio

“Min Box Area Ratio” is a floating-point input that compares each box’s area with total image pixels. When greater than 0, boxes at or below the threshold are filtered; setting it to 0 disables this filter. A threshold that is too high loses small text. Default: 0.

How the settings take effect

Detection branches and outputs

flowchart TD
    A["Input image"] --> B{"Import fixed YOLO boxes?"}
    B -->|Template and save text| C["Use imported boxes and build raw mask"]
    B -->|No or condition not met| D["Main detector: size/rearrangement/thresholds/unclip"]
    D --> E{"YOLO OBB enabled?"}
    E -->|No| F["Main boxes + detection mask"]
    E -->|Yes| G["YOLO auxiliary boxes: conf"]
    G --> H["Merge by overlap/containment/area"]
    H --> I{"SFX filter enabled?"}
    I -->|Yes| J["Keep YOLO-supported or bubble-exempt boxes"]
    I -->|No| F
    J --> F
    C --> K["text regions -> OCR"]
    F --> K

Long-image rearrangement tiles the image for detection and maps coordinates back during preprocessing. It can increase detection time and memory use. Here mask_raw means the raw mask returned by the detector or built from imported boxes; later stages decide how OCR and mask refinement consume it.

Imported labels and hybrid detection

flowchart LR
    A["Same-name annotation in yolo_labels/"] --> B["Read YOLO boxes"]
    B --> C{"Template + save text?"}
    C -->|Yes| D["Skip main detection"]
    C -->|No| E["Main detector + YOLO OBB"]
    E --> F{"Containment/overlap conditions"}
    F -->|Replacement| G["YOLO box replaces main box"]
    F -->|Below threshold| H["Append YOLO box"]
    F -->|Overlap without replacement| I["Remove duplicate YOLO box"]
    E --> J{"SFX filtering"}
    J -->|Bubble exemption| K["MangaLens bubble-mask exemption"]
    J -->|Include bubble text| L["Filter bubble text too"]

The diagrams express only branches confirmed by source; they are not a complete workflow diagram.

Interactions and caveats

• The selected offline detector model must load on the selected device; GPU/ONNX backend issues affect detection but hardware installation is outside this page.
• YOLO OBB and SFX filtering require the auxiliary YOLO model. Without auxiliary results, the main detector result remains the fallback.
• load_text disables YOLO OBB. Imported labels can bypass the main detector under the template/save-text condition; in the normal flow they replace detector boxes and preserve or build a raw mask where possible.
• text_threshold, box_threshold, unclip_ratio, and area ratio belong to different postprocessing layers. Broad thresholds pass noise to OCR; strict thresholds can miss text.
• Long-image rearrangement, larger detection sizes, and YOLO assistance consume more resources. For OOM or missing-model errors, reduce size or disable assistance rather than sharing real credentials.