Image Classification Architectures Summary

Summary

We learned all the popular image classification architectures and the novelties they brought us. Let’s do a quick review to sum up:

• AlexNet and VGG: These are architectures that use only standard convolutions with different kernel sizes.

• InceptionV1: It uses a network-in-network approach to make deeper networks in smaller sizes and with fewer model parameters. Additionally, InceptionV2 brings batch normalization usage to the convolutional layers.

• ResNet: It discovers residual blocks to handle vanishing gradient problems in deep neural networks.

• MobileNetV1: It uses depth-wise separable convolutions to tackle model size and parameter amount

• MobileNetV2: It uses inverted residual blocks with model size and parameter amount to tackle again.

• EfficientNet: It uses compound scaling to obtain better-scaled models while reducing the model size and parameter amount.

How to choose the model?

Choosing which model to use depends on our task. If we have limited memory or want our model size to be small, we could go for MobileNet or EfficientNet since we can play with its scaling until we find a good fit. If accuracy is our priority rather than model size, a deep ResNet architecture would be reasonable to use. Or, knowing all the fundamental components of a convolutional neural network, we could create our custom model according to our needs.