In this paper, a novel joint loss Generative Adversarial Networks (GAN) framework is proposed for thermography nondestructive testing named Defect-Detection Network (DeftectNet). A new joint loss function that incorporates both the modified GAN loss and penalty loss is proposed. The strategy enables the training process to be more stable and to significantly improve the detection rate. The obtained result shows that the proposed joint loss can better capture the salient features in order to improve the detection accuracy. In order to verify the effectiveness and robustness of the proposed method, experimental studies have been carried out for inner debond defects on both regular and irregular shaped carbon fiber reinforced polymer/plastic (CFRP) specimens. A comparison experiment has been undertaken to study the proposed method with other current state-of-the-art deep semantic segmentation algorithms. The promising results have been obtained where the performance of the proposed method can achieve end-to-end detection of defects.