Understanding the Roles of Geometric Forms and Proportions in CNN-Based Image Classification

The performance of Convolutional Neural Networks (CNNs) in image-based AI/ML applications can be affected by factors such as: number of training datasets used, geometric forms of subjects in images, and resizing for capturing geometric proportions. To the authors’ best knowledge, adequate information is not available on the influence of some of these factors on the accuracy of image classification using a CNN driven by deep learning techniques. To address the stated lacunae, two custom datasets were designed: one that contains cars, trucks and roses to investigate the possible effect of geometric form variation, and a second case study consisting of the images of lionesses and kittens, both belonging to the cat family, to assess the impact of proportional resizing. To bestow confidence on the results obtained, the effects of training dataset size, and values of hyper-parameters on accuracy of image classification are also investigated. The results indicate that a natural form may be more difficult to classify as compared to man-made geometry, and maintaining geometric proportions can add to rate of success of image classification.