coreml-vision

CoreML / Vision 本地推理

模型管理

目录结构

Resources/
└── Models/
    ├── YOLOv8n_v2.mlpackage       # 版本号在文件名中
    ├── MobileNetV3.mlmodel
    └── CLIP_Quantized.mlpackage   # 量化版标注

加载规范

• 所有 .mlmodel / .mlpackage 文件放在 Resources/Models/ 目录
• 模型加载使用懒加载，禁止在 App 启动时同步加载
• 模型版本通过文件名区分（如 YOLOv8n_v2.mlpackage），禁止覆盖旧版本
• 启用 Metal 加速：

final class ModelManager {
    static let shared = ModelManager()

    private var yoloModel: VNCoreMLModel?

    func loadYOLO() throws -> VNCoreMLModel {
        if let model = yoloModel { return model }
        let config = MLModelConfiguration()
        config.computeUnits = .all  // CPU + GPU + Neural Engine
        let model = try YOLOv8n_v2(configuration: config)
        yoloModel = try VNCoreMLModel(for: model.model)
        yoloModel?.inputImageFeatureName = "image"
        yoloModel?.outputFeatureName = "var_894"
        return yoloModel!
    }

    func unloadAll() {
        yoloModel = nil
    }
}

---

Vision Pipeline — 静态图像

完整推理流程

final class ImageDetector {
    private let model: VNCoreMLModel

    init(model: VNCoreMLModel) {
        self.model = model
    }

    func detect(in image: UIImage, confidenceThreshold: Float = 0.5) throws -> [Detection] {
        guard let cgImage = image.cgImage else {
            throw VisionError.invalidImage
        }

        let request = VNCoreMLRequest(model: model) { request, error in
            // 结果在下方处理
        }
        request.imageCropAndScaleOption = .scaleFill

        let handler = VNImageRequestHandler(cgImage: cgImage, options: [:])
        try handler.perform([request])

        guard let observations = request.results as? [VNRecognizedObjectObservation] else {
            return []
        }

        return observations.compactMap { observation in
            guard let label = observation.labels.first,
                  label.confidence >= confidenceThreshold else { return nil }
            return Detection(
                label: label.identifier,
                confidence: label.confidence,
                boundingBox: observation.boundingBox
            )
        }
    }
}

struct Detection {
    let label: String
    let confidence: Float
    let boundingBox: CGRect
}