Aiming at the great difficulties in collecting fabric data-sets and defecting fabric detection, an algorithm of fabric defect detection using deep learning combined with traditional methods is proposed in this paper. Firstly, an autocoding network based feature pyramid structure is proposed, which only needs normal samples for learning. Secondly, in the detection phase, multi-model fusion at the same scale is proposed to reduce the false alarm rate and remove the interference of texture noise. The experimental results show that the learning method proposed in this paper has a detection rate of over 98% for linear defects and over 84% for planar defects. It has more application value in practice.