In nuclear power plants, it is a common procedure to use video inspection when reloading fuel assemblies regularly. Because of the turbulence generated by residual heat of nuclear fuel assemblies, a video sequence suffers from severe geometric deformations and is hard to be used to position the location hole of fuel assemblies. The paper proposes a novel algorithm to recover a geometrically correct image of nuclear fuel assemblies or scene from a video sequence distorted by turbulence underwater and achieve the precise position of the location hole. At the first, an average image is utilized to compare with image sequences to estimate nonrigid registration based on B-splines. By the estimated nonrigid registration, new image sequences are obtained and are used to get a new average image. After multiple iterations, the better image sequence can be shown. Then the better image sequence are divided into image patches. The more blurry and severely distorted image patches are removed in image patch sequences. Finally, the selected image patches are synthesized together. Based on restored images, template matching is utilized to quickly find the initial position of the location hole. And then the sub-pixel centroid method is used to achieve the sub-pixel position in the image. The calibrated camera parameters are utilized to calculate the position of the location hole of the fuel assemblies. Experiments verify that the algorithm can online locate the center of location holes on recovered images underwater, and has high measurement precision.
When underwater camera is used to carry out the visual inspection after fuel reloading in nuclear power plants, heat exchange between fuel assemblies and water can generate underwater turbulence, which causes imaging distortion. Turbulence severely affects core verification of nuclear fuel assemblies, serial number of which should be identified. With the aim to recover the images from a video sequence severely distorted by turbulence, an image enhancement method is proposed. At first, an image quality assessment metric FSIM is used to select the better quality frames. Next an iterative robust registration algorithm is used to eliminate most geometric deformations and recover the water surface. The temporal mean of the sequence is utilized to overcome the structured turbulence of the waves through the algorithm. Finally, the sparse errors are extracted from the sequence through rank minimization to remove unstructured sparse noise. After image processing, optical character recognition is performed by KNN and CNN, obtaining high recognition rates of 99.33%, 100% respectively. The experimental results show that the suggested method significantly performs better in distorted image restoration and image text recognition on the task of visual inspection of nuclear fuel assemblies.