Lightweight Colorized Deep Learning Model Of Near-infrared Image With Fusion Layer

Near infrared image colorization is one of the key technologies of intelligent detection system.Many important night vision or low illumination scenes,such as mines,wild animal observation points,military bases,etc.,need use near the infrared image to achieve comprehensive monitoring.The advantage of the near infrared image is that can reflect the thermal radiation information of the target scene and near infrared image is not sensitive to the change of scene brightness.Near infrared image belongs to gray image,but do not conform to the human visual habits,and make it impossible for us to observe the scene large gray area details,so you need to color in order to increase its color and texture information,and to reduce the operator's visual fatigue,enhance the capacity of observers to the scene of the judgment of the situation.Because the current image colorization method based on color migration method rely on the reference image or color transmission method based on the color line,those method are extremely dependent on human intervention,and the process of artificial find the reference images and color the lines is very cumbersome,ser:iously affect the speed of the whole process of coloring,unable to adapt to the requirement of intelligent detecting system for adaptive,so this article use deep learning model to complete the color of the near infrared image.this paper needs 3000-5000 images that similar to near infrared image scene as image colorization network module of the requirements of the training set,this paper uses the lightweight joint characteristics of image recognition as near infrared image recognition network,in the near infrared image of object recognition accuracy,and greatly shortens the time of training and testing of equipment hardware is reduced requirements;Then,pass the identifying network,identified the scene and target categories contained in the near-infrared image to be colored,and the image sets with similar scenes and targets are selected in the ImageNet data set as the training set of the color-coded network.In view of the traditional network(CNN)--the details of image colorization effect is poor,easy to appear problem,add the Inception global V4 classifier to extracted the global features of the near infrared image,and fused with the local features that obtained by downsampling,get the fusion feature vector,then input local features to the decoder network to complete the color of the image and size of the recovery,finally,output the image that after the colorization.After testing,it is compared and analyzed with the existing image colorization methods,The results show that the color images obtained by this method are more detailed and have clearer edges.