Article: Image fusion using a transfer learning-based convolutional neural network Journal: International Journal of Information and Communication Technology (IJICT) 2025 Vol.26 No.3 pp.72 - 88 Abstract: Convolution neural network (CNN) is a deep learning model that is widely used in image recognition, image classification. However, traditional deep learning models require extensive annotation information during training, leading to prolonged training times that directly impacts efficiency and performance. In order to solve this problem, this paper proposes a five-layer convolution neural network structure based on transfer learning to extract, train and fuse features of source images. First, an improved VGG-19 network is used to extract the preliminary features of the source images, and the training samples are transferred to the encoder for deep feature extraction by setting the VGG-19 network parameters. Then, the extracted feature samples and a five-layer U-Net neural network construct the decoder for feature reconstruction. Batch normalisation is applied to prevent over-fitting of the model. Finally, the loss function is applied layer-by-layer in supervised learning to obtain the quadratic decision graphs that are used to fuse the source images to generate the output images. The proposed model in this paper demonstrates a significant enhancement in the visual effect of images compared with other models. Inderscience Publishers - linking academia, business and industry through research

Title: Image fusion using a transfer learning-based convolutional neural network

Authors: Mudan Lv; Aiping Cai

Addresses: School of Information Engineering, Jiangxi University of Technology, Nanchang City, 330098, China ' School of Information Engineering, Jiangxi University of Technology, Nanchang City, 330098, China

Abstract: Convolution neural network (CNN) is a deep learning model that is widely used in image recognition, image classification. However, traditional deep learning models require extensive annotation information during training, leading to prolonged training times that directly impacts efficiency and performance. In order to solve this problem, this paper proposes a five-layer convolution neural network structure based on transfer learning to extract, train and fuse features of source images. First, an improved VGG-19 network is used to extract the preliminary features of the source images, and the training samples are transferred to the encoder for deep feature extraction by setting the VGG-19 network parameters. Then, the extracted feature samples and a five-layer U-Net neural network construct the decoder for feature reconstruction. Batch normalisation is applied to prevent over-fitting of the model. Finally, the loss function is applied layer-by-layer in supervised learning to obtain the quadratic decision graphs that are used to fuse the source images to generate the output images. The proposed model in this paper demonstrates a significant enhancement in the visual effect of images compared with other models.

Keywords: deep learning; convolutional neural network; CNN; transfer learning; image fusion.

DOI: 10.1504/IJICT.2025.144458

International Journal of Information and Communication Technology, 2025 Vol.26 No.3, pp.72 - 88

Received: 28 Aug 2024
Accepted: 10 Dec 2024
Published online: 13 Feb 2025 *

Title: Image fusion using a transfer learning-based convolutional neural network

Keep up-to-date