| Summary: | The resolution of a digital core image obtained by CT scanning is inversely proportional to the core size. The sample size of the core plunger is large; therefore, the core image has low-resolution and loses microstructure information, which is difficult to meet the needs of follow-up research on core pores, fractures, rock skeleton and so on. By selecting a suitable position on the plunger sample, drilling one or more small-size subsamples for scanning to obtain the high-resolution image, the plunger core subsample image has both high-resolution and low-resolution. Due to the limited core image data of small-size plug samples, it cannot meet the demand for the number of training samples in super-resolution reconstruction based on depth learning. To solve the above problems, this paper first uses the low-resolution plunger sample image and its down sampled image to form the training data set, and then uses the SRResNet algorithm for model pretraining. Secondly, all the feature extraction layer parameters in the pre-trained model are frozen, and the model is retrained with the measured high-resolution plunger sample images and the corresponding low-resolution image in the plunger samples. The experimental results show that after 100 epochs of training, the loss function value of the SRResNet algorithm based on transfer learning is close to 0, and the PSNR value reaches 34. And the algorithm converges very fast, after about 30 epochs of training, the PSNR value is close to 34. Based on the transfer learning and SRResNet algorithm, the details such as small pores and textures that cannot be clearly displayed in the original image can be reconstructed, which can effectively realize the super-resolution reconstruction of digital core images.
|
|---|