Research On Infrared Image Semantic Segmentation Technology Based On Deep Learning

Infrared imaging is based on thermal radiation of objects.Infrared imaging is widely applied in the military,civil and other fields,for it is not affected by the light,has a large detection distance and strong anti-interference ability.In the application of automatic driving,security monitoring and quality inspection,it is necessary to understand the scene of infrared image by computer vision method,and realize the analysis of the content of the captured images.Scene understanding produces various computer vision tasks,such as image segmentation,scene classification,object recognition and detection,semantic segmentation,etc.For the infrared image reflects the difference of the thermal radiation between the objects and the background,which has no color information,and other inherent characteristics,such as low contrast,blurred edges and less information to express the texture,it is difficult to understand the whole infrared scene automatically by the traditional computer vision algorithms.In recent years,deep learning has achieve several state of the art results in computer vision tasks."Semantic segmentation" is a basic method to realize the image scene understanding by labeling the images with pixel-level semantic labels.It solves two sub tasks as image segmentation and classification.Adopting deep convolution neural network for supervised learning,semantic segmentation on visible image data sets has made a breakthrough progress.But there is still insufficient relevant research on infrared images.Therefore,this thesis address the task of sematic Infrared image semantic segmentation with deep learning.The main contributions of this thesis include the following aspects:1)Plan and build an infrared semantic segmentation dataset,which offer basic conditions to research of relative algorithms.Dataset plays a very important role in the research of deep learning algorithms.At present,there is a lack of open data set for the research of infrared image semantic segmentation.In this thesis,we set up the data set for the semantic segmentation task with outdoor scene image.The images mainly include 4 classes of objects,and are manually labeled.Data augmentation method is used to further expand the dataset to make up for the lack of data.The data set provides 1000 original 14 bit images,10 thousand times of augmented images and image preprocessing after 8 bit image enhancement,the corresponding labels,training and test set spilt rules and code for calculating accuracy of model.2)Combine the primary idea of ResNet and Inception Net,a 100 layer parallel residual network PresNet – 100 is put forward,which can be used to classification and semantic segmentation of infrared images.The PresNet-100 network consists of 4 sets of parallel residual structures with different sizes of the 16 filters.In order to ensure the diversity of the network potential sub network,reduce the depth of the hierarchy and increase the network width.Compared with ResNet-101,PresNet-100 reduces the network parameters and improves the training speed.Training and verifying PresNet on ImageNet dataset.It is proved that PresNet-100 has faster computation speed than ResNet-101,and the convergence speed is faster and the feature expression ability is better.3)Build a semantic segmentation network named Multi-PresNet with PresNet-100.Using fully convolutional framework,atrous convolution,multi-scale network measures and PreNet-100 to build semantic segmentation network.Compare it with three other network based on VGG-16.The training and testing experiments are carried out on the infrared datasets.Adopting the "pre-training on visible light image dataset and finetuning on infrared image dataset" method to train the 4 networks.The initial parameters of the networks are obtained from pre-training networks on Pascal VOC datasets and Cityscapes datasets.Then test the trained models on infrared testing datasets.This method can make up the lack of images of infrared image dataset.Experiments show that the semantic segmentation network based on multi-scale PresNet has higher prediction accuracy and good network performance.4)A high dimensional conditional random field refine algorithm based on superpixel is proposed.Post-processing the output scores of semantic segmentation network,which refines the inaccurate contours of objects in semantic segmentation resluts.Add two layers of superpixels to CRF model,extracting separately from the linear compressed 8-bit images and pre-processed enhanced 8-bit images.The higher order conditional random field model is expressed as the sum of three parts: unary pixel potential,pairwise pixels potential and super-pixels potential.Compared with the original fully connected CRF algorithm,the SLIC+CRF algorithm can better present the homogeneous patch information and get more accurate contour.Meanwhile,research on post-processing algorithms of semantic segmentation based on edge predicting and domain transform algorithm,which further improves the edge accuracy of semantic segmentation outputs.