Because the traditional detection algorithm is not accurate enough to detect cotton packaging defects and the recognition rate of small target defects is not high enough, an improved Faster R-CNN deep learning network is proposed to detect five defects such as damage, stain, hole, impurity and thread end in cotton packaging. Image enhancement is realized by preprocessing the image, then the RPN and ROI structure in Faster R-CNN are improved. In order to strengthen the detection ability of small target defects, the feature pyramid network structure is fused in the backbone network, and finally the ROI is bilinear interpolated to solve the problem of pixel deviation caused by multiple quantization. Experiments show that the average accuracy of the improved network for cotton packaging surface defect detection is 91. 34%, which is 9. 08% higher than the traditional algorithm.