Infrared light and visible light image fusion method based on residual dense network and gradient loss

Abstract

The invention discloses an infrared light and visible light image fusion method based on a residual dense network and gradient loss, and the method comprises the steps: putting an infrared light image and a visible light image into a pre-trained VGG-16 network, carrying out parameter extraction, and carrying out standardization and normalization of the extracted feature image values, and obtaining a weight block; putting the infrared light image and the visible light image into a network model for end-to-end unsupervised learning to generate a fused image model; training the fusion image model, in the training process, using a loss function for continuously conducting back propagation iteration, updating the learning content, wherein the loss function is based on the weight block optimization gradient; Fusion can be carried out on the existing public infrared light and visible light image data set, the edge information of the infrared light image and the texture details of the visible light image can be effectively extracted from the fused result, and the edge information of the infrared light image and the texture details of the visible light image are combined to form a fused image better conforming to the visual effect of human eyes.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to an infrared and visible light image fusion method based on residual dense network and gradient loss. Background technique [0002] Image fusion plays a pivotal role in computer vision, which extracts key information from different input images, synthesizes them together, and outputs a fused image with better visual effects. Infrared light and visible light image fusion is an important branch of image fusion. The infrared sensor can monitor the object around the clock, and generate an infrared image through the thermal radiation emitted by the object, and has excellent performance on the outline of the object in various environments. However, infrared images have poor texture details and low resolution. Visible light images have rich texture information and high spatial resolution, which can reflect the real environment of the detected object and adapt to the human visu...

AI Technical Summary

Problems solved by technology

These methods can accurately extract feature information, but require complex matrix operations, and the problem of edge blur still exists

In recent research, methods based on deep learning: DeepFuse, SESF Fuse and Attention FGAN can improve the shortcomings of traditional methods, but there are also some limitations

First, deep learning networks usually extract feature maps directly from the previous convolutional layer, ignoring global information, resulting in low-quality fusion results

Second, some methods do not employ an end-to-end model, but an encoder-decoder model with a fusion strategy

Too simple a fusion mechanism may make the edges of the image unclear

Finally, the design of the loss function will also affect the results of network training

If an appropriate loss function is not chosen for network training, the value of the loss function may converge slower and the training speed may decrease

Images