Augmenting for the best and most efficient image classification model.
You can find the complete report here: https://drive.google.com/file/d/15UOpucwdgoAsiyRUbjOJuQsKPSU0yyGs/view
In this section, we discuss the architecture of the Defensive Adversarial Mix-Up (DAM) Neural Network and the individual components that make this technique novel. Image filters, image Mix-Up, and Defensive Adversarial Mix-up are the three main components of the DAMN strategy. The below figure depicts the DAM technique’s modular architecture.
Image filters are the first component of the model and is divided into two sub-modules: attention filtering and environ-mental filtering. The aim of attention filters is to highlight the important sections of the input image, such as the targetobject, while fading out the surrounding regions. Environmentalfilters, in contrast to attention filters, aim to augment several surrounding scenarios for the target object, and it is performed in three variants: occlusion scenarios, road conditions, and weather conditions. The goal of filtering is to aid the neural network model in determining the object’s importance in relation to its surroundings. In the image Mix-up portion of our model we perform four different types of mix-ups after acquiring the filtered images: Linear Mix-Up, Vertical Concatenations, and Horizontal Concatenations. The algorithms for performing the mix-ups are discussed in detail in the paper, and the strategies used for defensive adversarial training for each of the mix-ups are also detailed.