ResNet50V2 is a convolutional neural network architecture renowned for its deep layers and skip connections, which facilitate effective training and learning of complex features. As part of our final model development, we have integrated ResNet50V2 as the backbone architecture, leveraging its pre-trained weights to extract hierarchical features from input images.

Our model utilizes transfer learning, a powerful technique that enables the transfer of knowledge from pre-trained models to new tasks. By fine-tuning the parameters of the ResNet50V2 backbone, we adapt the model to our specific dataset, enhancing its ability to classify images accurately.

To augment the dataset and increase its diversity, we apply data augmentation techniques during training. This involves generating synthetic variations of the input images, such as rotations, flips, and shifts, thereby enriching the training data and improving the model's robustness against variations in real-world images.

In addition to data augmentation, we employ a reduced learning rate on plateau strategy to dynamically adjust the learning rate during training. This adaptive approach monitors the validation loss, reducing the learning rate when the model's performance plateaus, and increasing it when improvements are observed. By fine-tuning the learning rate, we ensure efficient convergence and prevent overfitting.

Complementing the ResNet50V2 backbone, we design a custom top classifier comprising dense layers, batch normalization, and dropout regularization. These components further refine the model's ability to capture intricate patterns in the data while preventing overfitting by introducing stochasticity during training.