Zero-tap document scanning for Android. Point your camera at a document — DocShot automatically detects the boundary, captures when stable, corrects perspective distortion, and saves a clean rectangular image. No corner dragging, no multi-step workflows.
Every phone has a document scanner buried 3 taps deep in the camera app, and it still asks you to adjust corners manually. I wanted something faster: open the app, point at a document, get a clean scan. The kind of tool I'd actually use daily instead of just knowing it exists somewhere in my phone's settings.
- Zero-tap auto-capture — Detects document stability across 20 consecutive frames, locks autofocus, then captures automatically
- Real-time preview — Live quadrilateral overlay tracking the detected document boundary at 30fps
- Perspective correction — Sub-pixel corner refinement + homography warp produces a clean rectangular output
- Orientation detection — Sobel gradient + ink-density analysis ensures the result is right-side up
- Aspect ratio adjustment — Defaults to A4, auto-snaps to common formats (A4, US Letter, ID card, business card) with manual slider override, verified via homography decomposition
- Aspect ratio lock — Lock your chosen ratio so it persists across successive scans of same-format documents
- Gallery import — Rectify existing photos with automatic detection or manual corner adjustment
- Manual corner adjustment — Draggable handles with magnifier loupe when auto-detection needs help
- Post-processing filters — B&W (adaptive threshold), High Contrast (CLAHE), Even Light (gradient correction for angled lighting)
- Flash control — Torch toggle for low-light scanning, persisted across sessions and across captures (re-enables automatically after each scan)
- Confidence-gated capture — High confidence goes straight to result; low confidence routes to manual adjustment; very low is suppressed entirely
- Multi-strategy detection — 5 preprocessing strategies (standard, CLAHE, saturation, bilateral, morphological) with scene analysis to handle varied backgrounds
Classical computer vision pipeline — no ML models, no cloud services:
- Scene analysis — Histogram-based assessment selects optimal preprocessing strategy
- Preprocessing — Adaptive Gaussian blur + strategy-specific enhancement (CLAHE, bilateral, morphological closing)
- Edge detection — Canny with automatic threshold selection from median pixel intensity
- Contour extraction — External contours approximated to polygons, scored by area, convexity, angle regularity, aspect ratio, and edge density
- Corner refinement — Sub-pixel corner detection (
cornerSubPix) on the best quadrilateral candidate - Homography & warp —
getPerspectiveTransform+warpPerspective(INTER_CUBIC) to a properly-sized output rectangle - Post-processing — Optional filtering (B&W, contrast, lighting correction) on the rectified result
- Language: Kotlin
- UI: Jetpack Compose + Material 3
- Camera: CameraX with Camera2 interop (autofocus control, intrinsics extraction)
- Computer Vision: OpenCV Android SDK 4.12
- Build: Gradle Kotlin DSL, AGP 8.7
- Min SDK: 24 (Android 7.0) / Target SDK: 34
# Clone
git clone https://github.com/Artemarius/DocShot.git
cd DocShot
# Build debug APK
./gradlew assembleDebug
# Install on connected device
./gradlew installDebug
# Run tests
./gradlew testDebugUnitTest
# Build signed release AAB (requires keystore.properties — see keystore.properties.template)
./gradlew bundleReleaseRequirements:
- Android Studio Hedgehog (2023.1.1) or later
- Android SDK 34 + NDK
- OpenCV Android SDK 4.12 (set
OPENCV_ANDROID_SDKenvironment variable) - Physical device recommended (camera features don't work in emulator)
app/src/main/java/com/docshot/
├── ui/ # Compose screens + navigation
│ ├── CameraScreen # Camera preview, quad overlay, auto-capture UX
│ ├── ResultScreen # Before/after view, filters, aspect ratio, save/share
│ ├── CornerAdjustScreen # Manual corner handles + magnifier loupe
│ └── SettingsScreen # Preferences (auto-capture, haptics, format, filters)
├── camera/ # CameraX setup, frame analysis, AF control
├── cv/ # Document detection, rectification, post-processing
│ ├── DocumentDetector # Multi-strategy preprocessing + contour pipeline
│ ├── QuadRanker # Candidate scoring + confidence computation
│ ├── QuadSmoother # Temporal smoothing + stability detection
│ ├── Rectifier # Homography + warp + aspect ratio adjustment
│ └── PostProcessor # Filters (B&W, contrast, even light)
└── util/ # Permissions, image I/O, gallery save
| Metric | Target | Measured (Galaxy S21) |
|---|---|---|
| Detection latency | < 30ms/frame | ~15ms typical |
| Full capture pipeline | < 200ms | ~120ms |
| APK size (arm64) | < 30MB | 26MB |
| Memory usage | < 150MB | ~100MB typical |
74 tests total: 47 unit tests + 27 instrumented tests covering:
- Corner ordering and quad scoring
- Confidence computation and edge density validation
- Aspect ratio scoring against known document formats
- Contour analysis (partial documents, small documents)
- Synthetic image regression tests (low contrast, shadows, colored backgrounds, noise, overexposure)
DocShot processes everything on-device. No internet connection, no accounts, no analytics, no data collection. Privacy Policy
MIT