HDR Image Reconstruction And Recognition Technology Based On Deep Learning

Oil is one of the important energy sources to maintain the normal operation of modern society.Once the oil leakage occurs in the process of oil exploitation,it will cause seri-ous ecological disaster and huge loss of resources.Video monitoring technology in recent years,much attention in the oil field safety inspection,because of the video surveillance image image has the characteristics of intuitive and convenient,the oil security inspection on-line monitoring the introduction of computer vision technology,timely find possible faults in the oil field production process,can save the consumption of human resources and to ensure the quality and efficiency of security inspection.LDR(Low Dynamic Range Image)Image recognition technology is used in tradi-tional oil leak detection to obtain LDR images of oil production equipment and judge oil leakage incidents.Although this technology has the characteristics of Low cost and simple use.However,the technology must be in a good climate and exposure conditions,still can achieve relatively ideal results,easy to be restricted by lighting conditions and weather conditions,robustness is not strong.The different shapes of oil spill areas,the complexity of oil field environment and the shadow of light and radiation are the great challenges in oil spill detection.In view of the above problems,this thesis analyzes and summarizes previous research work,proposes to use deep learn-based High Dynamic Range Image re-construction and recognition technology to monitor oil production equipment in oil fields,and performs fusion and reconstruction of LDR images with different exposure levels.It solves the problems of oil leak detection caused by overexposure,weak exposure and shadow,and provides the basis for oil leak detection and monitoring.The main research contents of this thesis are as follows:1.Collect and construct oil leakage data set.For the existing LDR-HDR image data resources are relatively rare,and there is no publicly available oil spill scene data set.To solve the above problems,this thesis integrates the previous data resources on HDR image reconstruction to simulate different oil spill scenes in the oil field and uses Canon camera EOS5D3 to shoot LDR-HDR image pair for the scenes.The obtained LDR-HDR image data are preprocessed,including image screening and expansion,etc.Finally,the reconstructed HDR images were annotated to construct the oil spill data set.2.Synthesize HDR images using multi-exposure LDR images.In this research,based on the UNET network model,the Density Block module was introduced to improve it into a multi-scale Dense UNET network to generate HDR images from multi-exposure LDR images.The network uses a coarse-to-thin method consisting of three subnetworks to gradually reconstruct HDR images.On the coarse scale,the network predicts the global information(such as color,context)of HDR images from LDR images.On the medium scale branch,the network outputs the middle level details by learning the neighborhood pixels.Low-scale branching is used to preserve the details of the LDR image and pre-dict the high-frequency information that is not captured by the original image.By thinning modules of multi-scale HDR images,the method of this research can generate HDR im-ages containing more details.In addition,artifact problem is easy to occur in the process of multi-exposure image fusion,which can be solved by introducing optical flow method to align pixels.In this thesis,different scale models and different advanced HDR image reconstruction models are compared and studied.The experimental results show that this method is effective in both qualitative and quantitative aspects of HDR images.3.Target recognition and detection are carried out on the HDR image of the oil spill area based on the synthetic HDR image.By improving the target detection method based on Retinanet network model,the oil spill detection technology based on cyclic training method is proposed in this research.In order to enhance the learning ability of the network model for difficult samples from oil leakage,the cyclic training method is used to mine difficult samples and reduce the missed detection in target detection.In order to reduce the false detection probability of the model in the oil spill area,negative sample data sets are introduced into the network model for training.At the same time,the comprehensive testing method is used to improve the accuracy of target recognition and detection.By comparing the results of the above several target recognition and detection methods,the performance indicators such as the accuracy rate,recall rate and recall rate of oil leakage image recognition and detection have been greatly improved.