The idea of combining a flexible fiber bundle with the microscopic imaging system provides the possibility of the cross-scale detection of defects and textures on large-scale complex components. However, the pixelization artifacts caused by the inter-core spacing of the fibers degrade the image quality and make it difficult to identify the micro-features. A high-resolution reconstruction strategy is proposed based on dictionary learning. By training the high- and low-resolution image pairs after image registration, a coupled dictionary is obtained. Then high-quality images are obtained from the trained dictionary. Experimental results demonstrate that the pixelization artifacts can be effectively addressed, and the resolution of the reconstructed images can be promoted by 1.8 times.