以编程方式，我的代码检测两类图像之间的差异，并始终拒绝一个类，同时始终允许另一个类。
我还没有发现产生错误的图像和没有产生错误的图像之间有任何区别。但是必须有一些区别，因为产生错误的那些在 100% 的时间内这样做，而其他的在 100% 的时间内按预期工作。
特别是，我检查了颜色格式：两组中的RGB;大小：没有显着差异;数据类型：两组的uint8;像素值的大小：两组相似。
下面是两个永远不起作用的图像，然后是两个始终有效的图像：
此映像永远不起作用：https://www.colourbox.com/preview/11906131-maple-tree-and-grass-silhouette.jpg
此映像永远不起作用：http://feldmanphoto.com/wp-content/uploads/awe-inspiring-house-clipart-black-and-white-disney-coloring-pages-big-clipartxtras-illistration-background-housewives-bouncy.jpeg
此映像始终有效：http://www.spacedesign.us/wp-content/uploads/landscape-with-old-tree-and-grass-over-white-background-black-and-black-and-white-trees.jpg
此图像始终有效：http://www.modernhouse.co/wp-content/uploads/2017/07/1024px-RoseSeidlerHouseSulmanPrize.jpg
如何发现差异？
场景是我正在使用Firebase和Swift iOS前端将这些图像发送到Google Cloud ML引擎托管的convnet。有些图像一直有效，而某些其他图像则无法像上面那样工作。此外，当我使用 gcloud 版本预测 CLI 时，所有图像都可以工作。对我来说，问题必然是图像中的问题。因此，我在这里为成像组发布。为了完整性，代码会根据要求包含在内。
索引.js文件的代码包括：
'use strict';
const functions = require('firebase-functions');
const gcs = require('@google-cloud/storage');
const admin = require('firebase-admin');
const exec = require('child_process').exec;
const path = require('path');
const fs = require('fs');
const google = require('googleapis');
const sizeOf = require('image-size');
admin.initializeApp(functions.config().firebase);
const db = admin.firestore();
const rtdb = admin.database();
const dbRef = rtdb.ref();
function cmlePredict(b64img) {
return new Promise((resolve, reject) => {
google.auth.getApplicationDefault(function (err, authClient) {
if (err) {
reject(err);
}
if (authClient.createScopedRequired && authClient.createScopedRequired()) {
authClient = authClient.createScoped([
'https://www.googleapis.com/auth/cloud-platform'
]);
}
var ml = google.ml({
version: 'v1'
});
const params = {
auth: authClient,
name: 'projects/myproject-18865/models/my_model',
resource: {
instances: [
{
"image_bytes": {
"b64": b64img
}
}
]
}
};
ml.projects.predict(params, (err, result) => {
if (err) {
reject(err);
} else {
resolve(result);
}
});
});
});
}
function resizeImg(filepath) {
return new Promise((resolve, reject) => {
exec(`convert ${filepath} -resize 224x ${filepath}`, (err) => {
if (err) {
console.error('Failed to resize image', err);
reject(err);
} else {
console.log('resized image successfully');
resolve(filepath);
}
});
});
}
exports.runPrediction = functions.storage.object().onChange((event) => {
fs.rmdir('./tmp/', (err) => {
if (err) {
console.log('error deleting tmp/ dir');
}
});
const object = event.data;
const fileBucket = object.bucket;
const filePath = object.name;
const bucket = gcs().bucket(fileBucket);
const fileName = path.basename(filePath);
const file = bucket.file(filePath);
if (filePath.startsWith('images/')) {
const destination = '/tmp/' + fileName;
console.log('got a new image', filePath);
return file.download({
destination: destination
}).then(() => {
if(sizeOf(destination).width > 224) {
console.log('scaling image down...');
return resizeImg(destination);
} else {
return destination;
}
}).then(() => {
console.log('base64 encoding image...');
let bitmap = fs.readFileSync(destination);
return new Buffer(bitmap).toString('base64');
}).then((b64string) => {
console.log('sending image to CMLE...');
return cmlePredict(b64string);
}).then((result) => {
console.log(`results just returned and is: ${result}`);
let predict_proba = result.predictions[0]
const res_pred_val = Object.keys(predict_proba).map(k => predict_proba[k])
const res_val = Object.keys(result).map(k => result[k])
const class_proba = [1-res_pred_val,res_pred_val]
const opera_proba_init = 1-res_pred_val
const capitol_proba_init = res_pred_val-0
// convert fraction double to percentage int
let opera_proba = (Math.floor((opera_proba_init.toFixed(2))*100))|0
let capitol_proba = (Math.floor((capitol_proba_init.toFixed(2))*100))|0
let feature_list = ["houses", "trees"]
let outlinedImgPath = '';
let imageRef = db.collection('predicted_images').doc(filePath.slice(7));
outlinedImgPath = `outlined_img/${filePath.slice(7)}`;
imageRef.set({
image_path: outlinedImgPath,
opera_proba: opera_proba,
capitol_proba: capitol_proba
});
let predRef = dbRef.child("prediction_categories");
let arrayRef = dbRef.child("prediction_array");
predRef.set({
opera_proba: opera_proba,
capitol_proba: capitol_proba,
});
arrayRef.set({first: {
array_proba: [opera_proba,capitol_proba],
brief_description: ["a","b"],
more_details: ["aaaa","bbbb"],
feature_list: feature_list},
zummy1: "",
zummy2: ""});
return bucket.upload(destination, {destination: outlinedImgPath});
});
} else {
return 'not a new image';
}
});
问题是坏图像是灰度的，而不是我的模型预期的RGB。我最初首先通过查看形状来检查这一点。但是"坏"图像有 3 个颜色通道，这 3 个通道中的每一个都存储相同的数字---所以我的模型拒绝接受它们。此外，正如预期的那样，与我最初认为我观察到的相反，事实证明 gcloud ML 引擎预测 CLI 实际上也失败了这些图像。我花了 2 天时间才弄清楚这一点！