Classifying Human Thermal Images Using Deep Learning Technique in Artificial Intelligence

Abstract

Abstract Deep learning is a sub-field of machine learning that uses multi-layered artificial neural networks which has become very popular in recent times to perform tasks such as image and speech recognition, natural language processing, decision making and so on. It has several applications in mechanical engineering and especially in the manufacturing field, too. Applying image recognition concepts to inspect the defects in products (Quality control), scheduling maintenance of equipment based on the sensor data (Predictive maintenance), applying deep learning concepts to train the movements of robotic arms timely for welding & assembly in production workshops (Robotics), and using deep learning to optimize schedules & logistics (Supply chain optimization) could be some of the applications worth to mention. This paper focuses on classifying the thermal images of objects, especially thermal ones. This work discusses the application of three pre-trained convolutional neural network (CNN) models namely VGG16, MobileNetV2, and InceptionResNetV2 for classification of thermal images of two classes (two human faces) belonging to the same family with more resemblance in their appearance so as to increase the complexity in classifying. The procedure had two phases: The first phase has the training and testing with thermal images whereas the second phase has increased number of images with image augmentation technique to repeat the same process to improve the performance the models. The data-set used in the models consists of thermal images of two slightly different faces of the same family that were taken using a thermal imaging camera (Flir-E6). In the first phase, the data-set was divided into training, validation, and testing sets with each set consisting of 50 images, 30 images, and 20 images respectively. In the second phase, after the image augmentation, the data-set was divided into 350 training images, 120 validation images, and 50 testing images. The performance of the models was observed for 10,20, and 30 epochs. The learning rate was assigned to 0.0001 with Adam optimizer and a batch size of 10. The results showed that the model VGG16, in the initial phase, yields training and validation accuracy of 100% and 90% respectively. In the second phase, it yields training and validation accuracy of 100% and 90.83% respectively. The model could classify 19 out of 20 images correctly before image augmentation and 49 out of 50 images correctly after the image augmentation was applied. It shows an increase of 3% from 95% to 98% before and after image augmentation was applied. For the given input images, the other models, though they are well-established ones, did not perform better than VGG16.The confusion matrix was plotted for all the models for important results.