Image Processors

ID: uniflow.plugin.imageprocessors • Category: Media & AI • Version: v1.5.0 • Min Uniflow Version: Uniflow ≥ v1.4.0

Image Processors Plugin Reference Manual

1. Overview

Plugin Name: Image Processors

Type: Media & Analytics

Identifier: uniflow.plugin.imageprocessors

Platform: CrossPlatform (Windows & Linux x64/ARM64)

Description

Cross-platform computer vision and scene analytics plugin for multi-class object detection, privacy redaction, anonymization, and license plate Optical Character Recognition (OCR). The plugin ingests image file paths or byte arrays and performs high-speed automatic multi-class detection (humans/pedestrians, cars, trucks, buses, motorcycles, bicycles), privacy redaction (blurring human faces and vehicle license plates), as well as OCR character recognition and confidence scoring for detected plates.

2. Technical Architecture

The Image Processors plugin uses a high-performance cross-platform stack combining Microsoft.ML.OnnxRuntime for AI inference and SixLabors.ImageSharp for image rendering and redaction.

Built-in Embedded ONNX AI Models

All AI models are compiled directly into the plugin assembly as embedded resources. No external model folders or separate assets need to be deployed or distributed:

Model Resource Name	Neural Architecture	Input Tensor Shape	Description
`face_detection.onnx` Human Face Detection
↳	Ultra-Light Face Detector / YuNet	`[1, 3, 640, 640]` NCHW	Detects human facial bounding boxes for privacy anonymization.
`plate_detection.onnx` License Plate Detection
↳	YOLOv8 Vehicle Plate Detector	`[1, 3, 640, 480]` NCHW	Locates license plate regions across multi-scale resolution passes.
`plate_ocr.onnx` Character Recognition (OCR)
↳	LPRNet / CRNN Recognizer	`[1, 3, 24, 94]` NCHW	Recognizes vehicle license plate alphanumeric characters (`0-9`, `A-Z`) via CTC greedy decoding.
`object_detection.onnx` Multi-Class Object Detection
↳	YOLOv8 / YOLOv11 COCO Detector	`[1, 3, 640, 640]` NCHW	Detects multi-class objects: humans/pedestrians, cars, trucks, buses, motorcycles, and bicycles with bounding box coordinates.

Technical Characteristics & Custom Model Overrides

Direct Assembly Stream Loading: At startup, models are streamed directly from assembly memory into the ONNX runtime inference session.

Custom Model Overrides: Administrators can override any embedded model by setting the Custom Model Folder configuration setting or placing custom-trained face_detection.onnx, plate_detection.onnx, plate_ocr.onnx, or object_detection.onnx files in custom path defined in plugin configuration.

System Interaction & Exposed Catalog Routes

The Image Processors plugin integrates computer vision pipelines. It captures video frames from RTSP streams or image files, performing object detection, face/plate blurring, OCR text extraction, and scene analytics.

Architecture Diagram

VISUAL ARCHITECTURE FLOW DIAGRAM

Rendering Flow Architecture Diagram...

3. Configuration Parameters

The following configuration settings are available in the User Interface when creating or configuring an instance of this plugin source:

Configuration Setting	Description	Default Value
Default Redaction Style	Redaction mode: `GaussianBlur`, `Pixelate`, or `SolidColor`.	`GaussianBlur`
Blur Radius	Radius of Gaussian blur filter applied over detected bounding boxes.	`20`
Pixel Size	Tile block size when using `Pixelate` mode.	`15`
Confidence Threshold	Minimum AI detection confidence threshold (0.10 to 1.00).	`0.25`
IoU Threshold	Non-Max Suppression (NMS) Intersection-over-Union threshold for box overlapping.	`0.45`
Multi-Scale Inference	Evaluates frames across multiple image scales for close-up and distant detections.	`true`
Process Faces	Enable human face detection pass.	`true`
Process License Plates	Enable vehicle license plate detection pass.	`true`
Process OCR	Enable character recognition on detected license plate regions.	`true`
Process Objects	Enable general COCO object detection pass.	`false`
Process Vehicles	Enable vehicle classification pass (cars, trucks, buses, motorcycles, bicycles).	`false`
Process Persons	Enable human/pedestrian detection pass.	`false`
Output Image Format	Output container format (`Auto`, `Jpeg`, `Png`, `Webp`).	`Auto`
JPEG / WebP Quality	Compression quality percentage (1 to 100). Prevents file size inflation.	`80`
Custom Model Folder	Optional custom disk directory containing user-trained `.onnx` model files (`object_detection.onnx`, `face_detection.onnx`, `plate_detection.onnx`, `plate_ocr.onnx`).	(Empty - Embedded default)

4. Exposed Routes & Data Types

This plugin exposes catalog fields across Output Target (action sinks) and Event Input (trigger sources) rule graph nodes:

4.1 Output Target Action Nodes

The Output Target node acts as an action sink to perform object detection, vehicle classification, face/plate redaction, custom pipeline processing, and license plate OCR character recognition.

Target Name	Data Type	Associated Parameters	Parameter Type	Required	Description
Detect All Objects imageprocessor:detect_objects
↳	`String` / `Json`	`Image File Path` (`path`)	`String`	True	Detects all COCO objects (persons, vehicles, etc.) in an image.
Detect & Classify Vehicles imageprocessor:detect_vehicles
↳	`String` / `Json`	`Image File Path` (`path`)	`String`	True	Detects and categorizes vehicles (Cars, Trucks, Buses, Motorcycles, Bicycles).
Detect Humans / Pedestrians imageprocessor:detect_persons
↳	`String` / `Json`	`Image File Path` (`path`)	`String`	True	Detects human/pedestrian bounding boxes and counts.
Full Scene Intelligence imageprocessor:analyze_scene
↳	`String` / `Json`	`Image File Path` (`path`)	`String`	True	Runs combined full pass: Persons + Vehicles + Faces + Plates + OCR.
Blur Car License Plates imageprocessor:blur_plates
↳	`String` / `Json`	`Image File Path` (`path`)	`String`	True	Detects and blurs vehicle license plates in an image file.
Blur Human Faces imageprocessor:blur_faces
↳	`String` / `Json`	`Image File Path` (`path`)	`String`	True	Detects and blurs human faces in an image file.
Anonymize All (Faces & Plates) imageprocessor:anonymize_all
↳	`String` / `Json`	`Image File Path` (`path`)	`String`	True	Blurs both human faces and license plates in a single pass.
Read Car License Plate OCR imageprocessor:read_plate_ocr
↳	`String` / `Json`	`Image File Path` (`path`)	`String`	True	Detects license plates and runs OCR character recognition.
Custom Image Processing Pipeline imageprocessor:process_custom
↳	`Json`	Custom JSON payload	`Json`	True	Accepts advanced custom JSON configuration (`ImageProcessOptions`).

4.2 Event Input Trigger Nodes

The Event Input node triggers downstream rule execution whenever an image redaction, object detection, or custom processing operation completes.

Event Display Name	Family	Description
General Object Detection Completed event:detect_objects
↳	`EVENT`	Triggers when general COCO object detection finishes.
Vehicle Detection Completed event:detect_vehicles
↳	`EVENT`	Triggers when vehicle detection & classification finishes.
Person Detection Completed event:detect_persons
↳	`EVENT`	Triggers when human/pedestrian detection finishes.
Full Scene Intelligence Completed event:analyze_scene
↳	`EVENT`	Triggers when full scene analytics finishes.
Car License Plate Blur Completed event:blur_plates
↳	`EVENT`	Triggers when license plate redaction finishes.
Human Face Blur Completed event:blur_faces
↳	`EVENT`	Triggers when face redaction finishes.
Anonymize All Completed event:anonymize_all
↳	`EVENT`	Triggers when combined face and plate anonymization finishes.
License Plate OCR Read Completed event:ocr_read
↳	`EVENT`	Triggers when ALPR license plate OCR finishes.
Custom Processing Completed event:process_custom
↳	`EVENT`	Triggers when a custom processing pipeline run finishes.

4.3 Event Output Telemetry Fields

When an Event Input node triggers, it exposes the following output fields for downstream rule logic:

Field Name	Data Type	Description
`Success` Bool
↳		`true` if processing and file saving completed successfully without errors.
`Detected` Bool
↳		`true` if at least one object, face, or license plate was detected.
`TotalDetected` Int
↳		Total combined count of all detected entities in the image.
`TotalPersonsDetected` Int
↳		Number of humans/pedestrians detected.
`TotalVehiclesDetected` Int
↳		Total number of vehicles detected (Cars + Trucks + Buses + Motorcycles + Bicycles).
`TotalCarsDetected` Int
↳		Number of passenger cars detected.
`TotalTrucksDetected` Int
↳		Number of trucks/semi-trailers detected.
`TotalBusesDetected` Int
↳		Number of buses detected.
`TotalMotorcyclesDetected` Int
↳		Number of motorcycles detected.
`TotalBicyclesDetected` Int
↳		Number of bicycles detected.
`TotalFacesDetected` Int
↳		Number of human faces detected.
`TotalPlatesDetected` Int
↳		Number of license plates detected.
`FirstDetectedText` String
↳		Recognized alphanumeric plate text (e.g. `"AG77WPF"`) from the first detected plate.
`FirstTextProbability` Float
↳		Model confidence probability (0.0 to 1.0) for the recognized OCR text.
`InputFilePath` String
↳		Disk path of the original input image file.
`OutputFilePath` String
↳		Disk path of the saved redacted output image file.
`ExecutionTimeMs` Int
↳		Total execution duration in milliseconds (AI inference + rendering + encoding).
`ErrorMessage` String
↳		Detailed error description if `Success` is `false`.

5. Usage Examples

Scenario A: Automatic Camera Snapshot Redaction

Workflow Overview:

When an ONVIF IP camera detects motion (onvif.motion_alarm), Uniflow saves a JPEG snapshot to C:\Snapshots\camera1.jpg and then executes the Image Processors anonymize_all action to redact all human faces and vehicle license plates from the image. The event:anonymize_all completion event triggers a downstream FTP upload to archive the anonymized image to an offsite SFTP backup server.

Rule Node Configuration:

1. Event Input Node: ONVIF Motion Listener

Filter Event: onvif.motion_alarm

Exposed Fields: Topic (String), State (Bool)

2. Output Target Node A: ONVIF Action

Action Target: Save Snapshot to Folder

Destination Folder: C:\Snapshots\

File Name: camera1.jpg

3. Output Target Node B: Image Processors Action

Action Target: Anonymize All (Faces & Plates) (imageprocessor:anonymize_all)

Image File Path: C:\Snapshots\camera1.jpg

Logic Flow Diagram:

VISUAL ARCHITECTURE FLOW DIAGRAM

Rendering Flow Architecture Diagram...

Scenario B: License Plate OCR & Gate Trigger

Workflow Overview:

When a scheduled camera snapshot is captured, Uniflow executes the read_plate_ocr action to detect and recognize vehicle license plates using ONNX AI inference. The event:ocr_read completion event fires downstream logic that filters by FirstTextProbability > 0.85 and matches FirstDetectedText against an authorized vehicle list. If the plate matches, Uniflow writes a Modbus TCP coil (Coil 1 = True) to open the vehicle access gate barrier.

Rule Node Configuration:

1. Output Target Node: Image Processors Action

Action Target: Read Car License Plate OCR (imageprocessor:read_plate_ocr)

Image File Path: C:\Snapshots\gate_entry.jpg

2. Event Input Node: Image Processors Listener

Filter Event: event:ocr_read

Exposed Fields: FirstDetectedText (String), FirstTextProbability (Float), Success (Bool)

3. Logic Filter Node: And Condition

Expression: FirstTextProbability > 0.85 AND FirstDetectedText == "AG77WPF"

4. Output Target Node: Modbus Client Writer

Action Target: Direct Access

Type: Coils

Address: 1

Value: True

Logic Flow Diagram:

VISUAL ARCHITECTURE FLOW DIAGRAM

Rendering Flow Architecture Diagram...

Scenario C: Logistics Truck Gate Dispatch

Workflow Overview:

When a camera monitoring the freight gate captures a snapshot, Uniflow executes the detect_vehicles action to classify vehicles in the image. The event:detect_vehicles completion event fires downstream logic that evaluates TotalTrucksDetected > 0. If a truck is detected, Uniflow activates a Modbus coil (Coil 5 = True) to route the incoming freight vehicle to the designated loading bay, and sends an HTTP POST notification to the warehouse dispatch system.

Rule Node Configuration:

1. Output Target Node: Image Processors Action

Action Target: Detect & Classify Vehicles (imageprocessor:detect_vehicles)

Image File Path: C:\Snapshots\freight_gate.jpg

2. Event Input Node: Image Processors Listener

Filter Event: event:detect_vehicles

Exposed Fields: TotalTrucksDetected (Int), TotalVehiclesDetected (Int), Success (Bool)

3. Logic Filter Node: GreaterThan

Expression: TotalTrucksDetected > 0

4. Output Target Node A: Modbus Client Writer

Action Target: Direct Access

Type: Coils

Address: 5

Value: True

5. Output Target Node B: HTTP Client Action

Action Target: POST Request

URL: https://warehouse.local/api/dispatch

Logic Flow Diagram:

VISUAL ARCHITECTURE FLOW DIAGRAM

Rendering Flow Architecture Diagram...

Scenario D: Restricted Zone Intrusion Alarm

Workflow Overview:

When a perimeter security camera captures a scheduled snapshot of a restricted zone, Uniflow executes the detect_persons action to detect unauthorized human intrusion. The event:detect_persons completion event fires downstream logic that evaluates TotalPersonsDetected > 0. If humans are detected, Uniflow triggers a high-priority alarm in Avigilon Control Center and activates a physical siren via a Modbus coil (Coil 12 = True).

Rule Node Configuration:

1. Output Target Node: Image Processors Action

Action Target: Detect Humans / Pedestrians (imageprocessor:detect_persons)

Image File Path: C:\Snapshots\restricted_zone.jpg

2. Event Input Node: Image Processors Listener

Filter Event: event:detect_persons

Exposed Fields: TotalPersonsDetected (Int), Success (Bool)

3. Logic Filter Node: GreaterThan

Expression: TotalPersonsDetected > 0

4. Output Target Node A: Avigilon ACC Action

Action Target: Alarms (avigilon:alarm)

AlarmId: 901-ZONE-INTRUSION

Action: TRIGGER

Notes: Unauthorized human detected in restricted perimeter zone

5. Output Target Node B: Modbus Client Writer

Action Target: Direct Access

Type: Coils

Address: 12

Value: True

Logic Flow Diagram:

VISUAL ARCHITECTURE FLOW DIAGRAM

Rendering Flow Architecture Diagram...

Plugins Documentation