The invention relates to a Mura defect level judgment method and device based on deep learning. The method comprises the following steps that: setting a Mura defect level tag, and enabling a Mura defect image sample to correspond to the Mura defect level tag one by one according to a preset level; then, taking the Mura defect image sample as an input neuron to enter a network, carrying out training after feature extraction is carried out, and outputting a training result after the training result passes through a classifier, and obtaining a Mura defect image corresponding to the Mura defect level tag; and finally, outputting the Mura defect image test sample through feature extraction after the Mura defect image test sample passes through the classifier, and comparing with the Mura defectimage sample features with a Mura defect level tag to obtain the Mura defect level corresponding to classification. By use of the method, the Mura defect of a panel can be subjected to accurate and detailed judgment and output, the limitation of a traditional algorithm can be successfully evaded, and meanwhile, human cost can be greatly lowered.