2
$\begingroup$

The problem is based on binary classification (target and interference). I have an image dataset for which I can not use pixel intensities. I can use pixel coordinates only. Nevertheless, CNN will work very well on this dataset. However, I have built a model with logistic regression using two features. In this case, will my approach become very weak (according to any machine learning researcher) since I am not using CNN? The logistic regression-based approach is giving perfect results. Also, for error analysis, which metric would be appropriate in this approach?

$\endgroup$
2
  • $\begingroup$ How did you choose which two features to use from an image? And, what does it mean that you can use pixel coordinates and not intensities (do you mean pixel values)? $\endgroup$
    – gunes
    Commented Mar 28, 2020 at 17:58
  • $\begingroup$ Thanks for your comment. Usually, the image contains two or three separate objects. Since I can not use pixel intensities (i.e. pixel values), so I chose the first feature as the distance metric (distance from a fixed-line which will be present in the image all the time) and another as the size of the object. $\endgroup$
    – Salehin
    Commented Mar 28, 2020 at 18:17

1 Answer 1

1
$\begingroup$

If you're implementing the train/test (and validation if exists) correctly and getting good results (in which you should also be careful while evaluating your models as well, generally speaking), your method, even the simple logistic regression, is not weak. You don't need to use a CNN just because you have an image dataset.

A well-tuned CNN saves you from feature engineering, and finds the useful features itself. Apparently, you've found your useful features yourself, and so did what CNN would do in its very first layers, and fed these feature to the decision layer, namely the logistic regression in your case.

$\endgroup$
5
  • $\begingroup$ If the target and interference become very close, the distance metric won't work. In that case, which error metric should I focus on? Is the confusion matrix enough for error analysis? $\endgroup$
    – Salehin
    Commented Mar 28, 2020 at 18:43
  • $\begingroup$ but you said that the distance metric (actually distance from the line) is your feature, what's its relation to error metric? Yes, confusion matrix is good and much better than only accuracy. $\endgroup$
    – gunes
    Commented Mar 28, 2020 at 18:55
  • $\begingroup$ Usually, a pixel is categorized as interference if it is located very close to that line. But sometimes, the distance metric will yield very close values for target and interference. For example, it might give 1.01 units for target and 0.98 units for interference in a certain test set. The previous training datasets do not contain such samples. Then the target could be categorized as interference. For this particular scenario, how can I measure the error of the model? $\endgroup$
    – Salehin
    Commented Mar 28, 2020 at 19:18
  • $\begingroup$ so, correct me if I'm wrong: because one of your features isn't useful in some cases, you make errors on your binary classification, which I think is totally normal. Choice of your evaluation metric doesn't depend on your features. You said you're using confusion matrix, which is fine, (and also in the OP stated that log-reg is giving perfect results). So, what is the problem? $\endgroup$
    – gunes
    Commented Mar 28, 2020 at 20:40
  • $\begingroup$ If the confusion matrix is fine then I am good with the model. Thank you very much $\endgroup$
    – Salehin
    Commented Mar 28, 2020 at 23:43

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.