Research On Semantic Segmentation Of Infrared Images Based On Improved Deeplabv3+ And Ghostnetv2

In military and civilian fields,infrared imaging technology has been widely used due to its advantages such as long-range operation,immunity to light interference,and strong antiinterference ability.However,infrared images lack color information and have inherent features such as edge blurring,low contrast,and difficulty in expressing texture information.Therefore,traditional computer vision algorithms have difficulty comprehensively analyzing infrared scenes.Recently,deep learning has achieved significant success in multiple tasks of computer vision,and semantic segmentation is one of the basic methods for image scene understanding.By labeling pixel-level semantic information to the image,semantic segmentation can simultaneously achieve the two sub-tasks of image classification and segmentation.In recent years,deep convolutional neural networks have become the main method for solving semantic segmentation tasks in visible light images,but there is relatively little research on infrared images.this thesis adopts deep learning methods to conduct in-depth research on the semantic segmentation algorithm of infrared images.The specific research work and innovative points of this thesis are as follows:(1)Establishing and planning an infrared image dataset is critical for deep learning algorithm research,as the dataset serves as the foundation for training and testing deep learning models,thus improving their performance and effectiveness.In terms of research on infrared image semantic segmentation,the number of open datasets is still quite limited,making it essential to construct a dataset specifically for the automatic driving scene infrared image semantic segmentation task.Therefore,this thesis constructs dataset,which contains 8,000 image data mainly consisting of 8 common categories in automatic driving scenes,and has undergone meticulous manual annotation.This dataset will provide an important data resource for research in infrared image semantic segmentation and can be widely applied in fields such as infrared image analysis and understanding,and research and development of automatic driving systems.(2)In this thesis,we propose a Texture-Conditioned Convolutional Neural Network(TCCNN)for semantic segmentation of infrared images,which addresses the challenges of low resolution,weak texture,low contrast,and indistinct contours.The TC-CNN model incorporates a new Quantization and Counting Module(QCM)to statistically describe texture information,and an Enhanced Texture Module(ETM)to capture texture-related information,enhance texture details,and introduce spatial attention mechanisms to focus on useful information in the spatial domain.Furthermore,we propose a Texture Guidance Module(TGM)to adaptively guide the texture feature maps.Experimental results demonstrate that the proposed TC-CNN model with texture guidance outperforms existing methods in infrared image segmentation.(3)This thesis proposes a real-time semantic segmentation network for thermal infrared images based on improved GhostNetv2.Compared with traditional deep convolutional neural networks,this network significantly reduces the number of parameters and testing time while maintaining high segmentation accuracy,making it more suitable for low-storage and real-time requirements on mobile devices.The GhostNetv2 backbone network is improved by adding dilated convolutions to further improve segmentation accuracy while keeping the computational complexity unchanged.A new group-dilated spatial pyramid pooling module called the lightweight segmentation header structure(LSH)is proposed for multi-scale feature extraction and fusion,effectively reducing spatial information loss.In addition,an attention refinement module(ARM)is proposed to optimize the output features of each stage and further improve the network’s performance.Therefore,the lightweight GhostNetv2 network proposed in this thesis not only significantly reduces computational parameters but also extracts features with good discriminative performance,making it suitable for real-time semantic segmentation tasks for thermal infrared images.