Are the "artifacts" in select Keras MNIST training images really there or is my download corrupt?

Question

I'm having fun with a ludicrously well known and used dataset: mnist. I am doing it with a huge and well known tool: keras.

Please excuse the red dots, something else I was doing. I have otherwise not modified the image at all except via the process of plotting it to a file. All images below are plotted in R.

---

Image #48, a "5" digit: There is a perfect vertical slice through the upper segment. For all the letters, they look blurred. This slice is perfectly vertical and does not have blurring.

---

Image #54, a "4" digit: This has artifacts to the left and above the image. It looks like part of a border around the number.

---

Image #68, a "1" digit: This has artifacts to the left and above the image. It looks like part of a border around the number.

---

Maybe something is odd with my download or version, so perhaps someone else can verify this.

Are there the artifacts as shown in the images above, in the canonical dataset as presented by the well known tool?

Answer 1

There are many versions of MNIST digits now, and some of them are corrupted, binarized, or otherwise altered (see TensorFlow datasets) intentionally; but I don't think the keras MNIST digits has these "artifacts" intentionally added. These are real handwritten digits, so some of the "artifacts" are probably "people" mistakes. Some may be due to the fact that we are "zooming in" on low resolution images (consistent with your observation of images being "blurry"). I think the "stuff" around the digits that looks like a border may be adjacent text that was cropped out, but I really don't know.

Yes, I also get the same images as you do, so your download is not corrupt.