Research On Image Denoising And Object Detection In Nuclear Radiation Scenes

Due to the lethal effect of nuclear radiation on human body,the tasks in nuclear environment are usually completed by special robots instead of human.As an important data source of robot environment perception,image is the basis of a variety of advanced computer vision tasks.However,the high-energy particles in the radiation environment lead to the high degradation of the images,which affects the robot task.In order to ensure the flexibility of the robot,this paper studies the image denoising and object detection methods of nuclear radiation scene from the algorithm level.In image denoising,according to the characteristics of strong radiation scene noise image,a denoising network composed of noise learning unit and texture learning unit is designed,which has the characteristics of high denoising efficiency and rich texture details.In the noise learning unit,the specially designed backbone network extracts high-quality convolution features,and the additional receptive field blocks and spatial attention blocks improve the learning ability of the network;the texture learning unit learns the detailed texture features through a independent loss function.In order to further improve the comprehensive performance of the model,the Mish activation functions and asymmetric convolutions are used in the whole network.Compared with other common methods,this method achieves the highest PSNR and SSIM,which are 33.81 and 0.934,respectively.At the same time,the network’s FLOPs and parameters are much lower than those of comparing algorithms,so it is suitable for real-time performance on the embedded platforms of robots.In object detection,in order to ensure the balance of accuracy and speed of the algorithm in the embedded platforms,we make more improvements based on one-stage object detection algorithm SSD.Using VGG16 as the backbone network to reduce the size of the model;using Bi FPN as the feature fusion method to improve the learning ability of the network;using ATSS to obtain high-quality positive samples;using Focus Loss as the loss function to suppress sample imbalance.Compared with the original SSD,the proposed method has lower flops and higher FPS,which are 1.74 g and 94,respectively.At the same time,the average accuracy is only 1%,up to 88%.The algorithm in this paper is deployed in NVIDIA AXIVAR NT embedded platform,and the model transformation and algorithm cascade based on Tensor RT are completed.The results of the nuclear cold experiments dataset show that the PSNR after denoising is 35.5,the m AP of object detection is 0.874,and the overall FPS is 18.The introduction of image denoising improves the m AP of object detection by 36%.The results show that the proposed method is suitable for image denoising and object detection of nuclear environment robot in radiation scenes,indicating its high practical value.