ZeroLauncher/util/OnnxManager.cs

using Microsoft.ML.OnnxRuntime.Tensors;
using Microsoft.ML.OnnxRuntime;
using System.Diagnostics;
using System.Text;
using System.Drawing;

namespace Zerolauncher.util
{
	public class OnnxManager
	{
		public static OnnxVerify? onnxVerify;
	}
	public class OnnxVerify
	{
		private string _modelPath;
		private InferenceSession _session;
		private readonly string _inputName;
		static string[] labels = [
			"a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m",
			"n", "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z",
			"A", "B", "C", "D", "E", "F", "G", "H", "I", "J", "K", "L", "M",
			"N", "O", "P", "Q", "R", "S", "T", "U", "V", "W", "X", "Y", "Z",
			"0", "1", "2", "3", "4", "5", "6", "7", "8", "9"
		];

		public OnnxVerify(string modelPath = "plugin/model.bin")
		{
			_modelPath = modelPath;
			_session = new InferenceSession(_modelPath);
			// 获取模型的输入节点名称  
			_inputName = _session.InputMetadata.Keys.First();
		}

		public string RunInference(Bitmap image)
		{
			var tensor = CtcPreprocess(image, (3, 48, 320));
			var inputs = new List<NamedOnnxValue> { NamedOnnxValue.CreateFromTensor(_inputName, tensor) };
			using var results = _session.Run(inputs);

			var outputs = results.First().AsTensor<float>();
			var dimensions = outputs.Dimensions;

			var sb = new StringBuilder();
			float tmp; int tmp_index, last_index = -1;
			for (int j = 0; j < dimensions[1]; j++)
			{
				tmp = 0; tmp_index = 0;
				for (int k = 0; k < dimensions[2]; k++)
				{
					var tmp1 = outputs[0, j, k];
					if (tmp < tmp1) { tmp = tmp1; tmp_index = k; }
				}
				if (tmp_index == 0 || last_index == tmp_index)
				{
					// 试着过滤一下重复的字符，假设最大重复长度为2
					last_index = -1;
					continue;
				}
				last_index = tmp_index;
				sb.Append(labels[tmp_index - 1]);
			}

			return sb.ToString();
		}


		static DenseTensor<float> CtcPreprocess(Bitmap image, (int, int, int) recImageShape)
		{
			var (imgC, imgH, imgW) = recImageShape;
			var maxWhRatio = imgW / (float)imgH;
			var h = image.Height;
			var w = image.Width;
			var whRatio = w * 1.0f / h;
			maxWhRatio = Math.Max(maxWhRatio, whRatio);

			Debug.Assert(imgC == 3); // Assuming the image is in RGB format

			imgW = (int)(imgH * maxWhRatio);
			var ratio = w / (float)h;
			var resizedW = Math.Ceiling(imgH * ratio) > imgW ? imgW : (int)Math.Ceiling(imgH * ratio);

			var resizedImage = new Bitmap(image, new Size(resizedW, imgH));
			var paddingIm = new DenseTensor<float>([1, imgC, imgH, imgW]);
			for (int i = 0; i < resizedW; i++)
			{
				for (int j = 0; j < imgH; j++)
				{
					var pixel = resizedImage.GetPixel(i, j);
					paddingIm[0, 0, j, i] = (pixel.R / 255.0f - 0.5f) / 0.5f;
					paddingIm[0, 1, j, i] = (pixel.G / 255.0f - 0.5f) / 0.5f;
					paddingIm[0, 2, j, i] = (pixel.B / 255.0f - 0.5f) / 0.5f;
				}
			}

			return paddingIm;
		}

	}
}