2
$\begingroup$

I have a fairly general question pertaining to an image recognition ML model. I’ve recently developed an image recognition model using a single camera collecting more than 5000 images and then trained/developed the model.

My question is if I change the camera source (also changing resolution) would this impact the accuracy of the model? The scripts we’ve developed would still resize the image to the same size as the training set.

There is some discussion within our team about this and I’m hoping someone with a little more expertise in image recognition can add some insight.

Thanks!

$\endgroup$
1
  • 1
    $\begingroup$ Changing the camera resolution might affect the model's performance but in a specific case. Suppose the original camera took images of size 640 * 480. These images were resized to 256 * 256 for the model. Since a rectangular image was transformed into a square image, I assume that the image will be squished or stretched. Now, if the new camera takes images of size 512 * 512 and you resize them 256 * 256, there will be no squishing or stretching of features. This might a new data distribution for your model. Make sure you crop the necessary features and not resize the whole image. $\endgroup$ May 26, 2021 at 14:15

2 Answers 2

1
$\begingroup$

In general the accuracy will change (even dramatically as a google research team found out) if only with slight changes in the input image quality.

References:

  1. Understanding How Image Quality Affects Deep Neural Networks

Image quality is an important practical challenge that is often overlooked in the design of machine vision systems. Commonly, machine vision systems are trained and tested on high quality image datasets, yet in practical applications the input images can not be assumed to be of high quality. Recently, deep neural networks have obtained state-of-the-art performance on many machine vision tasks. In this paper we provide an evaluation of 4 state-of-the-art deep neural network models for image classification under quality distortions. We consider five types of quality distortions: blur, noise, contrast, JPEG, and JPEG2000 compression. We show that the existing networks are susceptible to these quality distortions, particularly to blur and noise. These results enable future work in developing deep neural networks that are more invariant to quality distortions.

  1. Does Deep Learning Have Deep Flaws?

A recent study of neural networks found that for every correctly classified image, one can generate an "adversarial", visually indistinguishable image that will be misclassified. This suggests potential deep flaws in all neural networks, including possibly a human brain.

$\endgroup$
0
$\begingroup$

I think by accuracy you mean validation/test accuracy. Any Deep Learning Model is highly affected by the distribution of training data. If there is a mismatch between train set and test/val set then your model will perform poorly. Now, what do i mean by distribution of training data? For example, you yourself captured 5000 images of puppies and cats to build a image classifier. Now you downloaded some puppies and cat picture from internet and try to predict them. It is very likely that your model won't perform well on this test because your test data is from totally different distribution. Changing camera/resolution won't affect your performance if you make sure your train and test data comes from same distribution(like your camera and very very similar cameras like your ones).

$\endgroup$

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.