Deploying High-Precision Document AI Under 14MB: Architectural Mechanics of Scan2PDF

Direct Answer: What Sets Scan2PDF's Document AI Apart?

Most consumer scanner apps transmit uncompressed document bitmaps over cellular interfaces to remote clusters running general-purpose segmentation APIs. Staksoft’s Document AI executes the entire vision pipeline directly on local mobile hardware using four dedicated, specialized neural nets (totalling ~14MB in aggregate). This achieves deterministic 12–25ms detection latency, zero byte leakage, complete offline resilience, and zero cloud server operating expense passed on to users.

Camera2 API (YUV_420_888 Frame Buffer) │ ├─► Downsampled Ring-Buffer (256x256x1 Grayscale) ──► Model 1: Quad Boundary Net [12-25ms] │ │ │ Calculates [p0, p1, p2, p3] Coordinates │ ▼ │ Capture Button Triggered / Auto-Stabilization Passed ────┘ ▼ Full-Resolution Sensor RAW/JPEG (e.g. 4000x3000) │ ├─► Stage 1: Homography Perspective Transform via Bilinear Interpolation │ ├─► Stage 2: Curvature Classifier Check (Fast 1.2MB Binary Gate) │ ├─► Flat Page Detected ────────► Bypass Dewarping (0ms overhead) │ └─► Book Spine Curve Detected ──► Model 2: 3D Surface Mesh Net [0.35-0.90s] │ ├─► Stage 3: Auto Color / Lighting Division (Model 3: Residual UNet [0.45s]) │ └─► Illuminant Field Extraction & Divisive Normalization │ ├─► Stage 4: Margin Inpainting (Model 4: In-Margin Finger Net [0.20s]) │ └─► Binary Mask Generation ──► Contextual Border Smoothing │ ▼ Clean Flattened RGB Document ──► Offline OCR Pipeline / Structured JSON Engine / Vector PDF

Architectural Specifications & Math

• Backbone: Depthwise separable inverted residual blocks derived from MobileNetV2 with width multiplier $\alpha = 0.35$.

• Quantized Parameter Size: 2,988,142 bytes (~2.98 MB).

• Loss Function: Combined Smooth $L_1$ Regression + Corner Heatmap Dice Loss:

Where $\hat{p}_k = (\hat{x}_k, \hat{y}_k)$ represents the normalized Cartesian coordinates for the four corners. Because regression heads alone can jitter across successive video frames, a Kalman state observer filters high-frequency frame tremor without inducing perceived temporal delay.

Surface Reconstruction & Backward Mapping

Rather than relying on iterative thin-plate spline meshes that drain battery reserves, Scan2PDF trains an encoder-decoder network to infer a dense 3D displacement vector field $\mathcal{D}(u, v) \rightarrow (x, y, z)$.

The model predicts local normal tangents $\mathbf{n}_i$, allowing Scan2PDF to perform backward grid unrolling in native C++ NDK, restoring curved lines to uniform horizontal alignments in under 0.90 seconds per page.

Physical Formulation of Retinex Decomposition

A photographed page $\mathbf{I}(x,y)$ in RGB color space is modeled as the element-wise product of surface reflectance $\mathbf{R}(x,y)$ and environmental illumination $\mathbf{L}(x,y)$:

Our 3.2 MB Auto Color network isolates the low-frequency illumination matrix $\hat{\mathbf{L}}(x,y)$ while preserving ink pigments inside $\mathbf{R}(x,y)$. Shadow-free normalization is then calculated directly:

Margin Inpainting & Strict Safety Fallbacks

Scan2PDF runs a lightweight U-Net segmenter generating a binary thumb mask $\mathbf{M}_{thumb} \in \{0, 1\}$. It enforces a strict border contiguity rule: any skin tone isolated within the document center is preserved (protecting passport photos and ID badges).

On held-out validation tests, the finger was successfully eliminated on 26 of 28 real phone photos, with false-positive triggers on only 2 of 104 fingerless pages.

Post-Training Integer Quantization (PTQ)

Quantizing weights from FP32 to signed INT8 shrank disk footprint by 74.2% (down to ~14MB aggregate) and sped up inference by 3.1x via vectorized ARM NEON SIMD acceleration.

Need Custom On-Device AI Models for Your Application?

Building low-latency, private, and quantized neural models requires specialized expertise in edge architecture, synthetic data simulation, and hardware acceleration. If your product requires custom computer vision, document intelligence, biometric segmentation, or offline neural OCR, Staksoft engineers custom on-device pipelines tailored specifically to your latency, size, and accuracy targets.

Can Staksoft build custom on-device AI models for my company?

Yes. Staksoft partners with enterprises, tech startups, and independent app publishers to design, train, compress, and deploy custom neural architectures directly into mobile apps. Whether you need real-time segmentation, offline OCR, or custom edge vision models, you can initiate a project discussion directly via our Contact Page.

Does Scan2PDF upload documents or photos to any cloud server?

No. All image scanning operations, boundary detection, spine flattening, Auto Color illumination normalization, finger erasure, and local AI OCR run entirely on the local device's CPU/GPU/NPU.

What are the minimum hardware specifications to run these models?

The models run on any device with Android 8.0 (API 26) or newer. The baseline latency benchmark of 12–25ms was measured on a Xiaomi Poco F1 equipped with a Qualcomm Snapdragon 845 SoC.

Are the model weights exposed or extractable on rooted devices?

All weights are encrypted using AES-256 with dynamic memory decryption routines baked into compiled C++ binary layers. Decryption keys are derived dynamically from verified runtime APK signatures.

Install Scan2PDF from Google Play Contact Engineering