Surface Image Analysis: Leveraging Convolutional Neural Network for Contact and Non-contact Imaging Methods

Abstract

This study aimed to compare the performance of elastomeric contact and non-contact object surfaces in CNN-based classification for pattern recognition. The study used capsule images of the rear surface of the fired cartridge cases, which have distinct traces depending on the gun they were fired from. The images were captured from the cartridge cases in three different categories: non-contact white light illumination, non-contact RGB-White light illumination, and elastomeric contact RGB-White light illumination. Elastomeric contact images were obtained from an elastomer with glass-integrated silicone rubber under RGB-White light. A general examination of the images from these three categories revealed that the elastomeric contact images displayed the lines and depth details on the capsule surface much better without surface glare compared to the other two image formats. Non-contact images under RGB-White light lacked a depth factor, but the surface lines were more visible, while under white light the surface was affected by glare and the surface lines were less pronounced. When these three categories of images were trained in CNN, it was observed that elastomeric contact images reconstructed the capsule surface and characterized the details on the surface much better, and as a result, they achieved higher accuracy in image classification.