0
$\begingroup$

I have images from deep sea, some are good quality and some barely anything is visible I want to classify the images (they're already labelled)

I performed few image enhancement tested it on few images (actual size) but when resize the images (244,244) the pixel look very visible (for example if thr image is gray the pixel of the object have some white pixels) Is resizing images important? Will the model learn even with the pixel issue

$\endgroup$
2
  • $\begingroup$ Out of curiousity, what is your dataset like? I've been approached about the possibility of deep learning on deep-water plankton tow videos and in a brief discussion was pretty discouraged due to their lack of labeled data. $\endgroup$ Nov 14, 2022 at 18:22
  • $\begingroup$ I have 20 videos each video is 5-7 minutes 1080 quality videos, i have 5 categories each category have either 3 or 4 videos. Each video is labelled $\endgroup$ Nov 14, 2022 at 18:39

3 Answers 3

1
$\begingroup$

Very abstractly, if you have "enough" information in the [channels, 244, 244] image, you should be able to train on it. If your images are large, downsizing them will remove information. If your images are small, upsizing them will introduce spurious information. Beyond that, you really need to work with your data to see if you have "enough" info for classification.

Often, with real-world classification, the target is small relative to the overall image. You might need to have an object-detection model detect a region of interest and then crop and resize that.

Blackwater imagery is challenging for a number of reasons:

  • Very high dynamic range between black background and highly reflective targets
  • Backscatter
  • Lots of filaments and fine structure in invertebrates (particularly challenging when it comes to the resizing question)
  • Very different shapes and classes than "ImageNet" or other common sets used for pretraining
  • Class imbalance
$\endgroup$
0
$\begingroup$

Upsampling as you mention, is usually for data augmentation, and it can help address the issue of class imbalance in the dataset.

However if you upsample too much, especially with already noisy data, the learning objective might get too difficult for the model and this will impact negatively fitting the model to the training set.

Instead, you can use masking or padding, which are techniques that do not deteriorate the original image.

Still, your main concern is to clean the noisy data, so I think training a denoising autoencoder, for example, might help improve the quality of the bad quality data.

$\endgroup$
0
$\begingroup$

There are various augmentation techniques available but according to your dataset you have to perfrom the up-sampling of the data as Adam Oudad said. You can perform Horizontal flip, Vertical flip operation.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.