A Semantic Segmentation Network For Infrared Scenes Based On Category Feature Enhancement

With the advent of the intelligent era,people's demand for intelligence of imaging products is increasing.For example,intelligent driving and unmanned security rely on intelligent products of vehicle-mounted and airborne imaging platforms.Among them,image segmentation technology is an important research direction for the realization of intelligent products.By analyzing the semantic information of the scene,segmenting the target and background in the scene,helping the landing and development of intelligent products has received extensive attention and attention.At the same time,people's demand for scene segmentation is not only in the visible light scene,but also in the infrared night scene.However,due to the high complexity of on-board and airborne scenes,there are many targets and it is difficult to distinguish them,and because of the characteristics of infrared imaging,the edges of targets are weak and the grayscale distribution is uneven,which makes it difficult to improve the segmentation accuracy of infrared on-board scenes.Therefore,this paper conducts the research on semantic segmentation technology for infrared images in vehicle and airborne scenes.The main research work is as follows:Infrared vehicle-mounted scenes are complex,and the details of infrared images are partially lost,and general image segmentation algorithms cannot meet the accuracy requirements of the task.Aiming at the classification error problem that easily occurs when segmenting the target edge,this paper designs a scene understanding network based on category prototype regression: CPRNet.The category prototype regression loss function is proposed in the network.By enhancing the feature correlation within the category,the distance within the category is shortened,thereby enhancing the image depth characteristics and improving the image details.At the same time,in the infrared vehicle-mounted scene,the spatial distribution of image details is also particularly critical.Therefore,CPRNet also designed a spatial and category attention mechanism based on category prototype regression.By weighting the depth features in the spatial dimension and the category feature dimension,the feature representation ability is improved,and the network is further improved in the infrared vehicle.The segmentation accuracy in the scene.Compared with the baseline network before the improvement,the accuracy rate is increased by 4.37%.The infrared airborne platform scene has the characteristics of small targets and weak edges.For this type of problem,we designed an image enhancement module based on super-resolution reconstruction as the input part of the network to perform size and edge details on the target in the image.Optimized and designed a frequency domain loss function to further optimize and improve the image enhancement module.The obtained image is then iteratively trained through our semantic segmentation network.At the same time,we designed an edge loss function for the semantic segmentation network to enhance the network's attention to the edge.The final experiment shows that the semantic segmentation network based on image enhancement is better than the general semantic segmentation network in the infrared airborne scene.Compared with the algorithm before the improvement,the accuracy rate has increased by 4.95%.