Contoso.AI.ImageSegmenterSINet

Contoso.AI.ImageSegmenterSINet

AI-powered image segmentation using Qualcomm's SINet (Salient Image Network) ONNX model. Separates foreground from background in images with NPU hardware acceleration.

Quick Start

using Contoso.AI;
using System.Drawing;

// Check if the feature is ready
var readyState = ImageSegmenterSINet.GetReadyState();

if (readyState != AIFeatureReadyState.Ready)
{
    // Prepare the feature (downloads QNN Execution Provider if needed)
    var readyResult = await ImageSegmenterSINet.EnsureReadyAsync();
    if (readyResult.Status != AIFeatureReadyResultState.Success)
    {
        Console.WriteLine($"Failed to initialize: {readyResult.ExtendedError?.Message}");
        return;
    }
}

// Create segmenter instance
using var segmenter = await ImageSegmenterSINet.CreateAsync();

// Segment an image
using var bitmap = new Bitmap("photo.jpg");
var result = segmenter.SegmentImage(bitmap);

// Extract foreground with transparent background
using var foreground = result.ExtractForeground(bitmap);
foreground.Save("foreground.png");

Features

• ✅ Foreground/background segmentation using SINet
• ✅ NPU hardware acceleration (QNN Execution Provider)
• ✅ Automatic model download at build time (both float and quantized models)
• ✅ Intelligent model selection: quantized for NPU, float for CPU
• ✅ Easy-to-use async factory pattern
• ✅ Multiple output formats (mask, overlay, extracted foreground)

Requirements

• Windows 10 SDK 19041+
• .NET 8.0+
• NPU recommended for best performance (falls back to CPU automatically)

Model Selection

The library automatically downloads two ONNX models at build time:

• Quantized (int8) model - Optimized for QNN NPU acceleration
• Float (fp32) model - Optimized for CPU execution

At runtime, CreateAsync() intelligently selects:

• Quantized model when QNN NPU is available
• Float model when falling back to CPU

This ensures optimal performance and compatibility across different hardware configurations.

API

License

MIT License - see LICENSE file for details.