Improvement of inspection accuracy by image composition using class activation map

Abstract

It is necessary to collect training data to inspect for foreign substances using AI. But to collect and annotate images of foreign substances mixed in normal products are costly. In this paper, training images of foreign substances mixed in normal products were composited from images of normal products and images of foreign substances. Inspection accuracy was evaluated for the case of training with actual images of foreign substances mixed in normal products and the case of training with composite images using CAM (Class Activation Map) . As a result, training with composite images improved inspection accuracy, while annotation cost was reduced.