Study on the Influence of Multiple Image Inputs of a Multi-View Fusion Neural Network Based on Grad-CAM and Masked Image Inputs

Neural network models are used successfully in many applications like traffic sign recognition in the automotive context, cancer detection in medicine engineering, machine monitoring in the manufacturing industry, et cetera. However, the decisions of a neural network model for a particular input sample in a classification task are mostly nontransparent. We propose techniques to determine which input image of a Multi-View Fusion Neural Network has the most influence on the prediction of the model for a particular image sample pair and which regions in the input images are important. In addition, a trained Multi-View Fusion Neural Network is studied regarding the question of influence. The results are convincing and show that the studied model learned similar concepts like a human.