How LADV Uses Deep Learning to Auto‑Generate Data Visualizations in Seconds

LADV is DataV's embedded intelligent design product that quickly learns and recognizes hand‑drawn sketches, infographics, and dashboard screenshots, automatically generating configurable visualizations within DataV.

1. Problem LADV Solves

It dramatically lowers the design cost of data visualization, allowing users to focus on requirement analysis, metric design, and data exploration rather than time‑consuming front‑end chart recreation.

Traditional visualization pipelines involve product, analysis, design, and front‑end teams, leading to slow, collaborative processes and high barriers for non‑designers.

Existing solutions like Tableau or Power BI provide templates that rarely match specific needs, forcing users to compromise on design quality and time.

2. Disrupting the Visualization Design Process

2.1 Traditional Workflow

Product and analysis teams gather requirements, designers create high‑fidelity mockups, and front‑end engineers implement them, causing delays and limiting non‑designers.

2.2 LADV‑Optimized Workflow

LADV reduces the workflow to a single step: users upload an image of a sketch or screenshot, and the system generates a ready‑to‑edit visualization in DataV, supporting further fine‑tuning.

3. Technical Solution

3.1 Chart Recognition Model

Using deep learning, LADV trains an object‑detection model on thousands of DataV templates to locate and classify chart types (bar, pie, etc.) and extract color schemes, a first in the chart‑recognition domain.

3.2 Color Recognition Model

The system extracts dominant colors from the source image, selects the largest‑area color as background, computes a WCAG‑compliant text color (contrast 7:1), filters similar hues, clusters remaining colors, and generates a harmonious palette for the dashboard.

3.3 Text and Font Recognition

Future releases will employ OCR and a ResNet‑18‑based font classifier to detect text content and font styles, restoring them in the generated visualization.

3.4 Chart Mapping

LADV maps recognized chart types to DataV component types, creates a basic JSON configuration with default styles and data, then refines colors and typography based on extracted information before rendering the interactive dashboard.

3.5 Framework Overview

The overall architecture combines chart detection, color extraction, text/font OCR, and mapping modules, orchestrated to produce a complete DataV dashboard configuration.

4. Project Demonstrations

Users can upload a template or hand‑drawn sketch; LADV generates an editable DataV dashboard, which can be further refined. Over 2,000 hand‑drawn designs collected with Zhejiang University’s CAD lab were used to train the model.

5. User Feedback

Tests with Power BI, Tableau, and generic dashboard examples showed high aesthetic scores. Even when baseline templates were less polished (e.g., Tableau Gallery), users preferred LADV‑generated visuals as starting points, indicating the tool’s ability to free designers from scratch work.

6. Future Outlook

After nearly a year of development and collaborations with academic partners such as Zhejiang University and TVCG editor Klaus Mueller, LADV aims to further integrate into all stages of visualization design, inspiring users to explore visual analytics more creatively.