Multi-label classification using Transfer Learning techniques

5 min readMar 2, 2023

COVID-19 is a contagious disease that caused thousands of deaths and infected millions worldwide. Using convolutional neural networks and transfer learning techniques we developed a highly accurate model to detect whether patients are infected with covid-19, some other virus, or nothing. This work mainly focuses on the use of CNN models for classifying chest X-ray images for coronavirus, viral infected, and normal patients

Techniques used

● TensorFlow
● Keras
● OpenCV
● VGG16
● VGG19
● ResNet-50

How does Convolutional Neural Network work?

A CNN consists of input, hidden, pooling, fully connected, and output layers. A CNN applies filters layers on the top of the input image to collect the important information carried by the pixels. Filter sizes can be selected as (3*3) or (5*5). Max pooling layer collects the highest pixel value from each filter generated. This process will continue and connecte as a one-dimensional array called a fully connected layer. Thus, a one-dimensional array can be as given input to ANN followed by hidden layers and output layers.

Developing Convolutional Neural Network

To develop an accurate CNN model without using fine-tuning techniques we applied 3 filter layers, and each filter has size (2*2), 3 max-pooling layers of size(2*2), and using strides(2*2).
To train the model on three different labels, we have 5000 images for each label and 2000 images for testing purposes.
The one-dimensional array acts as an input for ANN Architecture with 3 hidden layers with relu activation function and one output layer with a softmax activation function.

Performance metrics

What is Transfer learning CNN Techniques?

Transfer-learning techniques are neural networks trained on a 1000 labels dataset called ImageNet. Around 1.4 million images were used to develop CNN architectures. There are roughly 1.2 million training images, 50,000 validation images, and 150,000 testing images.

Transfer-learning technique VGG19

VGG19 has 24 different layers
* 16 convolutional layers
* 5 max-pooling layers
* 3 fully connected one-dimensional layers
* 1 output layer with a softmax activation function.

Architecture

• Fixed-size of (224 * 224) image was given as input to this network.
• Used kernels of (3 * 3) size with a stride size of 1 pixel, this enabled them to cover the whole notion of the image.
• Strides were used to apply the matrix on the top of filters and the input images.
• max-pooling was performed over 2 * 2-pixel windows with stride

Performance metrics

Transfer-learning Technique ResNet50

ResNet50 has 50 different layers with 48 convolutional layers, 1 max-pooling layer, and 1 Average pool layer. ResNet architectures are known as Residual networks with deep neural networks. ResNet architectures use a concept known as a skip connection. Using skip connection vanishing gradient problem reduced during the time of backpropagation.

Performance metrics

Transfer-Learning Technique VGG16

VGG16 has 21 different layers with 14 convolutional layers, 5 max-pooling layers and 3 hidden layers, and a softmax activation function at the output layer.
In transfer learning techniques we need to make sure that the pre-trained weights should not be changed and flatten layers from our data output itself. There is a padding of 1-pixel done after each convolution layer to prevent the spatial feature of the image.

Architecture

• Fixed-size of (224 * 224) image was given as input to this network.
• Used kernels of (3 * 3) size with a stride size of 1 pixel, this enabled them to cover the whole part of the image.
• max-pooling was performed over 2 * 2-pixel windows with stride.
• The final layer has 3 outcomes where we apply softmax and find the results.

Performance Metrics

Finally using fine-tuning VGG16, VGG19, and ResNet50 deep learning techniques in terms of extracting information from the data it was observed that VGG16 performed well in classifying covid, Normal, and viral pneumonia. However, excellence in high performance remained besides VGG16 with high precision and high model accuracy.