The present tile defect detection algorithms mainly rely on manual design features and classifier. In addition, they face debugging difficulties and insufficient robustness in practical applications. Therefore, we proposed a texture tile defect detection algorithm using the improved YOLOv3 model. First, a convolutional autoencoder was added in front of the Darknet-53; the reconstructed images with weak defects were fused with original images to get richer input information. Further, the K-means clustering method was used to get new and more suitable anchors. Finally, to solve the problem of insufficient samples, we used the weights of a pre-trained model trained on a common data set to initialize the network to improve convergence performance. Results show that the average accuracy of the improved model increased by 5 percent, besides it kept the prediction speed of the original model and could effectively detect texture tile holes and scratches.