An Industrial Application of Multi Target Detection in Thermal Images from Different Cameras with DeepLearning

Abstract

In this study, the main aim is to automatically perform the manual target detection process used in the camera field of view testing of mass-produced thermal cameras. A data set is prepared by taking images using different mass production cameras and different test systems. With this prepared data set multi target detection architecture is proposed. This proposed hybrid architecture consist of ResNet50 block, which is used for feature extraction, and YOLOv3 block. The accuracy of this proposed architecture to detect targets whose number and position changes in each image, compared with Minimum Output Sum of Squared Error(MOSSE), Single Shut Detection(SSD), Aggregate Channel Features(ACF), Recurrent Convolutional Neural Network(RCNN), FAST-RCNN, FASTER-RCNN, and YOLO versions target detection architectures. As a result of this comparison, it is seen that the proposed hybrid architecture has higher accuracy than other architectures. The use of proposed hybrid architecture in the camera field of view test of each camera produced with mass production will reduce the workload and increase the accuracy of the camera field of view calculation.