Year: 2024
Type: article
Abstract: The field of image classification using Convolutional Neural Networks (CNNs) to predict the principal object in an image has seen many recent innovations. One aspect that has not been extensively explored is the effect of the camera employed to acquire images for inference. We investigate this by capturing comparable images of five drinking vessels using six cameras in various scenarios. We examine the classification ranking of object classes when these images are input to an independently pretrained Resnet-18 model based on the ImageNet-1k dataset. We find that the camera used can affect the top prediction of object class, particularly in scenarios with a more complex background. This is the case even when the cameras have similar fields of view. We also introduce a metric called selectivity, defined as the mean absolute difference between prediction probabilities of similar relevant object classes (such as cups and mugs). We show that the effect of the camera is largest when the selectivity of the pretrained model between these object classes is small. The effect of camera choice is also demonstrated quantitatively by examining Cohen’s Kappa (κ) statistic. Finally, we make recommendations on mitigating the effect of the camera on image-classification inference.
Source: Applied Sciences
DOI: 10.3390/app15010246
Authors: Jason Brown, Andy Nguyen, Nawin Raj
Institution: University of Southern Queensland
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,728
OpenAlex ID: W4405921822