如何子类化Apple的VNImageBasedRequest以集成OpenCV处理？

要把OpenCV检测逻辑整合到Vision框架的请求流程中，你需要实现三个核心部分：自定义请求类、请求处理器，以及结果承载类。以下是完整示例：

1. 定义检测结果类（可选但推荐） #

创建一个类来承载OpenCV的检测结果，方便后续回调获取：

import Vision

class OpenCVDetectResult: VNRequestResult {
    let overlayImage: UIImage? // 可根据需求替换为检测坐标、目标数组等类型
    
    init(overlayImage: UIImage?, request: VNRequest) {
        self.overlayImage = overlayImage
        super.init(request: request)
    }
    
    required init?(coder: NSCoder) {
        fatalError("init(coder:) has not been implemented")
    }
}

2. 自定义VNImageBasedRequest子类 #

创建你的请求类，指定默认处理器并封装结果回调：

class OpenCVDetectRequest: VNImageBasedRequest {
    typealias CompletionHandler = (OpenCVDetectRequest, Error?) -> Void
    
    private(set) var result: OpenCVDetectResult?
    var completionHandler: CompletionHandler?
    
    override init(completionHandler: VNRequestCompletionHandler? = nil) {
        super.init(completionHandler: completionHandler)
        self.setDefaultHandler()
    }
    
    required init?(coder: NSCoder) {
        super.init(coder: coder)
        self.setDefaultHandler()
    }
    
    private func setDefaultHandler() {
        self.handler = OpenCVDetectRequestHandler()
    }
    
    override func results() -> [Any]? {
        guard let result = result else { return nil }
        return [result]
    }
}

3. 实现请求处理器 #

创建遵循VNRequestHandler的类，在这里封装OpenCV检测逻辑：

import CoreVideo
// 确保已导入OpenCV桥接库
import OpenCV

class OpenCVDetectRequestHandler: NSObject, VNRequestHandler {
    func perform(_ request: VNRequest, on image: VNImageBuffer, orientation: CGImagePropertyOrientation, options: [VNImageOption : Any]?) throws {
        guard let detectRequest = request as? OpenCVDetectRequest else {
            throw VNError(.invalidRequest)
        }
        
        // 将Vision的图像缓冲区转换为OpenCV可处理的格式
        var inputImage: UIImage?
        if let pixelBuffer = image.cvPixelBuffer {
            inputImage = CameraCaptureManager.GetImageFromPixelBuffer(pixelBuffer)
        } else if let cgImage = image.cgImage {
            inputImage = UIImage(cgImage: cgImage)
        }
        
        guard let validImage = inputImage else {
            throw VNError(.invalidImage)
        }
        
        // 执行你的OpenCV检测逻辑
        let overlayImage = opencv_wrapper.detect(in: validImage)
        
        // 赋值结果并触发回调
        detectRequest.result = OpenCVDetectResult(overlayImage: overlayImage, request: detectRequest)
        detectRequest.completionHandler?(detectRequest, nil)
    }
}

4. 整合到现有流程中 #

将自定义请求加入self.requests，和其他VNCoreMLRequest一起执行：

// 初始化自定义OpenCV请求
let openCVRequest = OpenCVDetectRequest { request, error in
    if let error = error {
        print("OpenCV检测出错：\(error)")
        return
    }
    
    // 获取并处理结果
    if let result = request.results?.first as? OpenCVDetectResult {
        if let overlay = result.overlayImage {
            DispatchQueue.main.async {
                // 主线程更新UI，比如显示叠加后的图片
                self.imageView.image = overlay
            }
        }
    }
}

// 添加到请求数组
self.requests.append(openCVRequest)

// 执行所有请求
let imageRequestHandler = VNImageRequestHandler(cvPixelBuffer: pixelBuffer, orientation: .up, options: [:])
do {
    try imageRequestHandler.perform(self.requests)
} catch {
    print("请求执行失败：\(error)")
}

注意事项 #

• 若OpenCV逻辑支持直接处理CVPixelBuffer，可跳过UIImage转换，直接使用image.cvPixelBuffer提升性能
• Vision请求默认在后台队列执行，更新UI必须切回主线程
• 可根据实际需求修改OpenCVDetectResult的内容，比如返回检测目标的坐标数组、关键点数据等

内容的提问来源于stack exchange，提问作者justin