Watermarkly

Automatically detects human faces in images using deep learning computer vision models (likely MTCNN, RetinaFace, or similar face detection architectures) and applies configurable blur filters to detected regions without manual selection. The system processes image tensors through a pre-trained neural network to identify face bounding boxes, then applies Gaussian or pixelation blur kernels to those regions in real-time or batch mode.

Extends face detection to identify and blur sensitive text regions (license plates, ID numbers, addresses, email addresses) using optical character recognition (OCR) combined with object detection. The system likely uses CRAFT or similar text detection models to locate text bounding boxes, optionally runs OCR to classify sensitive patterns (regex matching for phone numbers, license plate formats), and applies blur only to flagged regions.

Processes multiple images sequentially or in parallel through the detection and blur pipeline, likely using a job queue system (Redis, RabbitMQ, or similar) to distribute inference workloads across GPU/CPU resources. The system accepts a folder or file list, queues detection jobs, applies blur to each image, and returns a batch of processed images with progress tracking and error handling for failed detections.

Provides user-configurable blur parameters (Gaussian blur radius, pixelation block size, motion blur direction) and style presets (light, medium, heavy redaction) that are applied uniformly or selectively to detected regions. The system likely stores blur configuration as metadata or presets, allowing users to adjust blur strength before or after detection without re-running the detection model.

Provides a browser-based interface (likely React or Vue.js frontend) with drag-and-drop file upload, real-time preview of detected regions before blur application, and one-click download of processed images. The UI communicates with a backend API (REST or GraphQL) to submit images for processing and retrieve results, with progress indicators and error messages for failed detections.

Returns confidence scores for each detected region (face, text, license plate) indicating the model's certainty, allowing users to filter or review low-confidence detections before applying blur. The system likely provides a review interface where users can accept/reject individual detections, adjust bounding boxes, or manually add missed regions before finalizing blur application.

Exports blurred images in multiple formats (JPEG, PNG, WebP) with configurable compression levels and quality settings, preserving metadata (EXIF, color profile) or stripping it for privacy. The system likely uses image encoding libraries (libvips, ImageMagick, or native browser APIs) to transcode the blurred image tensor into the selected format with user-specified quality parameters.

Offers pre-configured redaction profiles (e.g., 'Legal Document', 'Healthcare Photo', 'Social Media Screenshot') that bundle detection sensitivity, blur strength, and export settings optimized for specific use cases. The system likely stores these as configuration templates that users can select before processing, with optional customization of individual parameters.