Which backbone model should I use for my Feature Classifier Model?

AI Discussions
1

Hi folks, a beginner to Deep Learning here and really hoping someone can shed some light on my questions. Here’s my problem;

I am building a Feature Classifier model. At the moment, I intend to use a ResNet backbone followed by a SoftMax layer. The model will be used to classify individual tree canopies into categories depending on how damaged the trees are. The input data I am using is fairly unique; this input data (i.e. training samples) are false images derived from photos taken from a drone. By “false image” I mean that these images consist of pixels that represent a single value, instead of the standard RGB images which may contain three separate R, G and B values. So based on this, I have some questions I am hoping someone could answer for me;

  1. As my input data are not regular photos, does this affect which backbone model I should use? It is my understanding that many of the popular ResNet models are based on ImageNet data, and of course this database is unlikely to contain false images of tree canopies similar to my input data. In other words, can a backbone model still be useful even if none of the photos it is trained on are similar to the input data I will be using?

  2. I am also curious to know, under which circumstances can it be better to not use a backbone model at all?

  3. It is my understanding that the ResNet models each contain a different number of layers i.e. ResNet 34 contains 34 neural layers, ResNet50 contains 50 layers and so forth. What are the advantages and disadvantages of using a backbone with more layers or a backbone with less?

Thank you for advance for any help. Even some insight to one of these questions would be really useful!

2

By “false image”, do you mean they are gray-scale instead of RGB?

3

For all intents and purposes, the images are grey-scale. (Technically, they are a per pixel calculation of NDVI).

4

I’m not sure entirely what you mean by “backbone model”.

Using a pretrained model (if one exists) is usually a faster method than training your own - as someone else has already gathered the training set and optimized the model architecture.

I’m not sure whether there is a pre-trained ResNet model for monochrome images.

The OpenCV package has a function for changing monochrome images into the equivalent RGB image.