Image Data Augmentation and Convolutional Feature Map Visualizations in Computer Vision Applications

Abstract

Deep neural networks (DNNs) perform exceptionally well in many vision tasks, including image classification, annotation, and object recognition. However, these networks are like a black box, and high-quality training datasets are required for deep learning models to achieve high performance. Due to the high cost of collecting a vast number of data samples, data augmentation techniques have been employed in many vision applications. Data augmentation aims to increase the dataset size without collecting new data while introducing variability. One of the means of augmenting the image data is by employing image transformations such as flipping, clipping, or rotation. Activation maps, also known as feature maps, illustrate how the filters are applied to the input image. The objective of visualizing a feature map for an input image is to comprehend what input features are captured in the feature maps. In this paper, we apply various transformations on images and investigate their effect on the multiple convolutional layers (at low, middle, and high levels) by employing intermediate feature map visualizations. We use the famous deep learning-based pre-trained network, VGG-16. Finally, we compare the visualization results of the image transformations at multiple levels and analyze their differences to evaluate the validity of these networks.