Defect detection plays an important role in yield improvement for TFT array process. Traditional manual recognition is inefficient, while the emerging convolutional neural network for target detection needs a lot of manpower in defect labeling. In order to realize the automatic detection of TFT array defects while reducing labor costs as much as possible, a TFT array defect detection method based on Generative Adversarial Networks and morphological reconstruction is proposed. In this method, the dataset used to train the network does not need to be manually labeled, which solves the problem of high manual labeling costs. This method first obtains the attention mask of the TFT array through the AttentionGAN network, secondly, selects the least significant pixel in the attention mask as the seed point, obtains the defect mark image and the defect mask image, and then performs the region growth for binary morphological reconstruction, finally get the bounding box of defect. This method can achieve an F1 score of 0.94 for the two-class classification of TFT array defects, which proposes a new idea for automatic defect detection of TFT array.