0

I'm working on Mask(Labels) generation for Semantic segmentation problem

I have a NumPy array of 250x321. It's a single-channel image. I've to annotate this image. But since this is a single-channel image, when trying to save it in the normal mode of matplotlib.pyplot, dimensions are changed along with a change in the channel (4). When storing just a single-channel image in png format, the image is all black, hence unable to annotate the image.

So I'm unable to annotate png format image (All Black) and also unable to use normal matplotlib.pyplot save function (Dim are changed along with channels)

Can someone suggest me how to achieve this job!

I've used code given below

plt.imshow(data[:, 400:650].T)
plt.tight_layout()
plt.xticks([])
plt.yticks([])
plt.axis('off')
plt.savefig('/content/new.png',bbox_inches='tight', pad_inches=0.0)

The above code is generating an array of shape 245x315x4 when used plt.imread when in the original array was of shape 250x321x1.

Quamer Nasim
  • 366
  • 1
  • 5
  • 18
  • 1
    It could be a few things for example a data type (are your images frloat or int) issue. However, it's difficult to help you without any code. In this particular case, you should create a minimum reproducible example (https://stackoverflow.com/help/minimal-reproducible-example). – Ash Jul 03 '20 at 11:10
  • @Ash, I've edited the question and added code, pls see – Quamer Nasim Jul 03 '20 at 11:58

1 Answers1

1

First of all, if you want to just save the image, avoid using plt.savefig and instead use either plt.imsave (or alternatives from other libraries, e.g. opencv's imwrite, or the tools from PIL, etc). The reason for this is that plt.savefig is implemented to save the entire figure, which means saving axes, labels, text etc. Furthermore, it takes a boolean argument Transparent which will handle the fourth channel. So, without knowing the inner workings and exact implementation of plt.savefig, it's a safe bet to say that it shouldn't be used to save grayscale images.

In your case however, since you're not using plt.savefig to plot anything besides the image itself, there is a simple workaround that I'll come to after covering two points in your question. First, your code is

plt.imshow(data[:, 400:650].T)

but since you want a grayscale image, you should set cmap="gray". Second, regarding this part

When storing just a single-channel image in png format, the image is all black

There isn't any info in your question on the range of values that are in your data, but I'm guessing that this is just an issue of "float" vs "uint8" datatype. Try setting data.astype(float64) or if using uint8, make sure you normalize it in [0,255] and not [0,1].

Now regarding the simple workaround, once you've figured out how to correct the above and saved your figure, you'll notice when loading the image with plt.imread that all your four channels have the same values (assuming you've set cmap=gray), so you can just select any of the four channels and discard the others.

But again, it's a better idea to use plt.imsave or opencv.

Ash
  • 4,611
  • 6
  • 27
  • 41