咨询Google Cloud Vision批量图像标注请求的计费规则与成本优化

我开发了一款从视频中提取大量帧的应用，目前使用Google Cloud Vision的Web检测功能，通过单个API请求处理每帧图像，运行状态良好，但希望找到更具成本效益的处理方案。考虑采用批量图像标注来优化性能，但存在疑问：一次批量图像标注HTTP请求是按1次事务计费，还是按批量中的图像数量计费？

当前使用的代码 #

exports.analyzemultipleFrames = async fileList => {
    let cloudVisionFrames = [];

    const filteredFramesToVision = fileList.filter((file, index) => {
        let filterFrame = false;
        if (index % appConstants.frameCountAnalyse == 0) {
            filterFrame = true;
        }
        return filterFrame;
    });
    const analyzeFramesPromise = await filteredFramesToVision.map(async file => {
        const imageAnalysisout = await analyzeFrame(file.fullPath);
        cloudVisionFrames.push(file.fileName);
        return imageAnalysisout;
    });

    const frameAnalysisResult = await Promise.all(analyzeFramesPromise);

    return {
        cloudVisionFrames: cloudVisionFrames,
        frameAnalysisResult: frameAnalysisResult
    };
};

exports.analyzeResubmittedFrames = async (framesArr) => {
    let cloudVisionFrames = [];
    const analyzeFramesPromise = await framesArr.map(async file => {
        const imageAnalysisout = await analyzeSingleFrame(file);
        cloudVisionFrames.push(file);
        return imageAnalysisout;
    });

    const frameAnalysisResult = await Promise.all(analyzeFramesPromise);

    return {
        cloudVisionFrames: cloudVisionFrames,
        frameAnalysisResult: frameAnalysisResult
    };
}

async function analyzeFrame(file) {
    // eslint-disable-next-line no-async-promise-executor
    return new Promise(async resolve => {
        fs.readFile(file, async (err, data) => {
            if (err) {
                resolve({ frameAnalysis: {}, frame: file });
            }
            let base64String = Buffer.from(data).toString("base64");
            let request = {
                image: { content: base64String },
                features: appConstants.cloudVisionFeatures
            };

            try {
                const [result] = await visionAnnotateClient.annotateImage(request);
                resolve({ frameAnalysis: result, frame: file });
            } catch (err) {
                logger("Vision Error" + err);
                resolve({ frameAnalysis: {}, frame: file });
            }
            //implement google cloud vision logic to getch files
        });
    });
}

async function analyzeSingleFrame(file) {
    // eslint-disable-next-line no-async-promise-executor
    return new Promise(async resolve => {
        let request = {
            image: { source: { imageUri: file } },
            features: appConstants.cloudVisionFeatures
        };

        try {
            const [result] = await visionAnnotateClient.annotateImage(request);
            resolve({ frameAnalysis: result, frame: file });
        } catch (err) {
            logger("Single Frame Vision ERROR" + err);
            resolve({ frameAnalysis: {}, frame: file });
        }
        //implement google cloud vision logic to getch files
    });
};

计费规则明确 #

Google Cloud Vision的批量图像标注请求并非按单次HTTP请求计费，而是按批量中每张图像的处理次数单独计费：

• 每一张图像的标注请求（无论是否包含在批量中）都会按照你指定的features类型（比如Web检测）单独计费。
• 批量请求只是将多个单图像请求合并为一个HTTP请求，减少网络往返次数、降低延迟，但计费总和与你单独发送这些请求的总费用一致。

优化建议 #

虽然批量请求不会直接降低计费成本，但能显著提升处理性能，结合以下策略可进一步优化整体成本与效率：

• 优化帧采样逻辑：你代码中已通过index % appConstants.frameCountAnalyse == 0过滤帧，可根据业务需求调整采样率，避免对冗余帧进行分析。
• 使用批量API替代单帧请求：替换当前的annotateImage调用为batchAnnotateImages，减少网络开销，提升并发处理效率。
• 优先使用Cloud Storage URI：批量请求中使用GCS存储的图像URI（如gs://bucket/path/frame.jpg），避免Base64编码带来的额外数据传输开销。

修改后的批量处理代码示例 #

exports.analyzemultipleFrames = async fileList => {
    let cloudVisionFrames = [];

    const filteredFramesToVision = fileList.filter((file, index) => {
        return index % appConstants.frameCountAnalyse === 0;
    });

    // 构建批量请求参数
    const requests = filteredFramesToVision.map(file => {
        cloudVisionFrames.push(file.fileName);
        // 假设文件已上传至Cloud Storage，使用URI；本地文件需先上传或保留base64逻辑
        return {
            image: { source: { imageUri: `gs://your-bucket/${file.fullPath}` } },
            features: appConstants.cloudVisionFeatures
        };
    });

    try {
        const [result] = await visionAnnotateClient.batchAnnotateImages({ requests });
        const frameAnalysisResult = result.responses.map((response, idx) => ({
            frameAnalysis: response,
            frame: filteredFramesToVision[idx].fullPath
        }));
        return { cloudVisionFrames, frameAnalysisResult };
    } catch (err) {
        logger("Vision Batch Error: " + err);
        // 批量请求失败时 fallback 到单帧处理
        const frameAnalysisResult = await Promise.all(
            filteredFramesToVision.map(file => analyzeFrame(file.fullPath))
        );
        return { cloudVisionFrames, frameAnalysisResult };
    }
};

内容的提问来源于stack exchange，提问作者Nishtha Bansal