Performance Evaluation of the Convolutional Neural Networks for Object Identification Using RGB and Binary Images

Convolutional Neural Network (CNN) is a topmost deep learning technique because of its capability to learn features autonomously with domain-specific images similar to the classical machine learning approaches. One common tactic for training CNN architecture is to transfer learned knowledge of a pre-trained network used to perform one task into a fresh task. The quest to detect objects more accurately remains an open research area. The use of traditional classification methods is widespread, but limited by time-ineffectiveness, and subjectiveness. The advent of CNN attempts to estimate and extract features inside images for the improved precision of image classification. This paper evaluates the two CNN models (CNN-1 and CNN-2) for object identification with two input image formats. The input image-types including: binary, and RGB colour images collected from the MINST and CIFAR-10 databases respectively. The training and validation of the two CNN models were conducted on the Google Colaboratory Virtual Machine with sampled images. The results showed that, the object identification task for the RGB colour images outperformed binary images-based system in terms of speed (14.20ms to 29.00ms). Conversely, the object identification task with the binary images was superior to the RGB colour images for the loss function (3.20% to 88.30%), and accuracy (99.10% to 71.06%). These results were comparable to existing studies using the speed and accuracy. Thus, the paper has demonstrated the image data types significantly impact on speed, loss, and accuracy, which offer guides for evolving reliable applications like surveillance, medical diagnosis, agriculture, and transportation systems.