Fused Image Classification using Pre-trained Deep Convolutional Neural Networks

Abstract

Image Quality Assessment (IQA) has got importance in the computer vision applications as it provides tool to evaluate and rate different image processing algorithms. Image Fusion is a process in which information from multiple images is combined into a single image. Due to specific nature of fused images present IQA methods have limitations for evaluation of Image Fusion algorithms. With the recent development of Deep Convolutional Neural Networks (Deep CNNs), No- reference image quality assessment is becomes reality. This article has proposed the pre-trained Deep CNNs based image fusion classification using Alexnet, VGG19, Inception V3 and ResNet-50. Four states–of–the-art image fusion algorithms used for image fusion are Laplacian Pyramid (LP), Shift Invariant DWT (SIDWT), Discrete Wavelet Transform (DWT) and Ratio Pyramid (RP). To achieve the effective IQA, sufficiently large dataset of synthetically fused images is created and same is evaluated by using Deep CNNs. The results show that recent deep CNN methods correctly classify the fused images into corresponding categories based on its fusion algorithms. The results are consistent with FR-IQA methods. ResNet-50 provides best classification accuracy with less number of epochs and time to converge due to sparse network connections.