Research On Object Detection Method Based On Visible And Infrared Image Fusion

The common data mode for object detection is visible light,but in the environment of weak light,dense smoke and fog,it is difficult to effectively complete the object detection task with a single mode of visible light image.Visible and infrared images are generated by different sensors,so the image features are different.Visible light images have rich details and background texture information,but are limited to producing high-quality images under good imaging conditions.In dense smoke and fog,and at night,visible light images may lose target information;Infrared images,based on their principle of thermal imaging,can retain target information with thermal radiation in harsh visual environments.The complementary spatial information between the two modalities of images can improve the effectiveness of object detection tasks.This article conducts research from two aspects: visible light image and infrared image fusion,and target detection based on fusion images.Firstly,this article studies the image fusion section and analyzes the advantages and disadvantages of traditional image fusion methods based on multi-scale transformations and deep learning image fusion methods based on convolutional neural networks.In response to the requirements of the project,in order to reduce the pixel position differences between different modal images,this study preprocesses infrared and visible light images to achieve the purpose of noise reduction and filling in missing information;In order to balance the fusion speed and fusion effect,a lower temporal and spatial complexity Laplace decomposition framework is adopted as the underlying decomposition framework.Combined with the idea of extracting image features using convolutional layers in deep learning,image morphology methods are used to preserve the target contour of infrared images and the background information of visible light images.The obtained information is used as weights to participate in Laplace decomposition and reconstruction simultaneously with multimodal images.The calculation results of entropy,edge gradient,and average gradient in the objective evaluation criteria of this method are 7.49,74.61,and 7.23,respectively,which are higher than other image fusion methods.Secondly,in the object detection section based on image fusion,this article analyzes the advantages and disadvantages of fusion at different stages of detection.In order to meet the flexibility of the combination of image fusion and object detection,this article adopts a method based on pre detection fusion.In network design,the fused image and low-frequency sub-band information generated during the fusion process are used as network inputs to increase detailed feature information.In the input method,two methods of feature pixel addition and connection are proposed.After experimental verification,the detection accuracy of these two methods is superior to the original network in three categories: human,vehicle,and aircraft.