Abstract: In order to identify fabric defects and reduce economic losses, aiming at the problems of low detection accuracy and insensitivity to small target detection in some existing network detection methods, this paper proposes a fabric defect algorithm ZS-Cascade RCNN based on cascade RCNN. In the feature extraction stage, deformable convolution was added to preserve the integrity of feature. The anchor frame was adjusted to meet the defect detection requirements of different aspect ratios to improve the detection effect. The cross and parallel ratio equalization sampling was used to equalize positive and negative samples. Experimental results show that the accuracy of ZS-Cascade RCNN algorithm is 4.5 percentage points higher and the average accuracy is 17.8 percentage points higher than that of the original algorithm. The effect of fabric defect detection is obviously improved.