训练和验证的准确性高，但甚至无法正确预测单个图像

Question

我已经在GTSRB（德国交通标志数据集）上训练了一个模型，它对43个类别的大约39K图像的大小从16x16到128x128不等。训练准确率= 97％，验证准确度= 91％。但是当我抓取自己的任何图像时，它永远无法正确识别。下面是我的代码图片

Libraries and accessing data is shown.i have also tried other way without using imagedatagenerator but same problem

here is my model. I also tried it with input shape(64,64) but same problem

here is how i am predicting my images

here is my model.fit results with loss and acc. %