When i start doing literature i observed that most of the authors used pre-trained CNN models for the identification and classification of plant leaf diseases. Why none of the authors concentrated to designed customised CNN model for this problem? is there any particular reason?
To make a robust CNN model it requires a large number of pre-labeled data sets. Which is quite difficult while tackling a practical problem. Moreover, the training of a model is a time-consuming process and requires a large storage space, which is quite costly in general.
You are right that a pre-trained model is highly customized with its parameters and size but the impact is low when you trained over a large number of data sets.
Srinivas Talasila
Deep neural networks such as Convolutional neural networks contain many different layers, and though this is not deliberate, each layer tends to extract features from the input data, with more general and simple features at the earlier layers such as lines, and more complex features at deeper layers such as shapes.
Traning a network that is able to do this well takes a lot of time, money, and labeled data.
If you have a small data set, it will be hard to create a good classifier, so what can you do?
You can take a pre-trained network that has feature extraction abilities on its lower layers, remove its last layers, and replace them with your own classification layers. In this way, you take advantage of the pre-trained layers that take a lot of time and data, to do the basic feature extraction of the images and only have to train for the remaining layers of the network to adjust it to do your task.
This is called "Transfer learning".
And the number of layers you are "reusing" is depending on your resources, time, computation power, the size of your data-set.
But using Pre-trained models parameters and size may be high compared to customized one. How can we overcome these problems.
The easy solution is to rescale your image to match the pre-trained network. This can be done by either expanding or shrinking your image, or zero-padding. However, if your input is too small, that expanding it will introduce too many artifacts or the image is very high resolution, that shrinking might cause it to lose important small detail, the only option is to change the input size of the network.
You can change the pre-trained network's input size by adding CNN layers before it and get to the correct image size before connecting to the pre-trained network by pooling and/or stride convolutions.
for more information, refer to this guide:
https://www.pyimagesearch.com/2019/06/24/change-input-shape-dimensions-for-fine-tuning-with-keras/
The claim of using transfer learning is not limited to plant disease, but in fact, it is applied in most applications nowadays. In general, designing a new architecture + hyper-parameter tuning need tremendous amount of time and resources.
You can train your own CNN, but you need enough training data. In addition, there are many details you must apply in the training process. transfer learning is a common approach in DL.
To make a robust CNN model it requires a large number of pre-labeled data sets. Which is quite difficult while tackling a practical problem. Moreover, the training of a model is a time-consuming process and requires a large storage space, which is quite costly in general.
You are right that a pre-trained model is highly customized with its parameters and size but the impact is low when you trained over a large number of data sets.
Srinivas Talasila
Pre-trained models are designed using lot of labeling data samples. So in most of image processing researchers utilized pre-trained models by performing transfer learning.
Usually, pre-trained CNNs have effective filters to extract information from the images because they are trained with a well-distributed dataset, and they have a good architecture. Basically, the filters in the convolutional layers are properly trained to extract the features of the images. If you train your custom CNN with class unbalance or with many similar images you can bias the training. Re-train a pre-trained CNN to tune it for your application is what is called transfer learning.
because it is trained million of epochs on millions of samples
It is a good practice to pick a pre-trained model and then fine tune parameters according to your area of application.
Also, there are similar features between the ImageNet dataset and plants which will help to perform well.
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing