JPEG Pleno compression can introduce artifacts in the compressed 4D light field images, which not only degrade their visual quality but also affect the performance of other image processing tasks. At present, there is a lack of research on JPEG Pleno light field coding distortion repair and JPEG Pleno coding distortion light field dataset for network training. In this paper, an end-to-end flexible blind convolutional neural network, namely JPLARNet, is proposed, which fully considers the characteristics of JPEG Pleno encoding and can predict the coding quality factor to control the trade-off between artifact removal and detail preservation. Specifically, JPLARNet initially employs a spatial-angle decoupling module to perform preliminary feature extraction on light field images. The extracted features are then processed by a multi-scale decoupler to obtain predicted compression factors and high-level semantic features. Subsequently, a compression factor fusion module is utilized to embed the predicted compression factors into the subsequent reconstructor module, and then embeds the predicted compression factors into the subsequent reconstruction module through the compression factor fusion module, thereby guiding artifact removal in compressed light field images. In addition, two modules, namely the LK Down-sample module and the mixed attention enhancement module, are constructed for downsampling and image enhancement of the reconstructed images, respectively. The experimental results show that on the six compression qualities of the constructed JPL-DATA dataset with compression factors ranging from 5000 to 200000, the average gain in YUV-PSNR/Y-SSIM of the light field image after artifact removal compared to before removal is 0.81 dB/0.025. By taking into account the characteristics of JPEG Pleno encoded light field image, the proposed method achieves light field images with higher subjective and objective quality than the JPEG artifact removal method of 2D images.