Abstract: Terahertz imaging has been one of the most promising technologies for non-contact detection in recent years. Among them, the internal quality of sunflower seeds can be detected using terahertz imaging. Specifically, terahertz waves can penetrate the shells of sunflower seeds, and then interact with their internal tissue structures and components, thereby obtaining their information of quality. However, the imaging speed has severely impeded the effective application of terahertz imaging in large-scale and rapid detection scenarios. Traditional terahertz imaging can usually consume a substantial amount of time to collect and then process the image data. Thus, it is highly required to accurately and rapidly detect the quality of sunflower seeds, in order to fully meet the high-efficiency requirements in industrial production or actual detection in modern agriculture. In this study, the compressed sensing with the Attention-Enhanced Super-Resolution Generative Adversarial Network (A-ESRGAN) model was proposed in the field of terahertz imaging. Rapid and accurate detection was also achieved in the plumpness of sunflower seeds. Furthermore, the crucial step was to select the measurement matrix in the initial stage during compressed sensing. Firstly, the Compressive Sampling Matching Pursuit (CoSaMP) reconstruction was adopted to verify the performance of different measurement matrices. According to the best overall performance, the Gauss matrix was then selected as the final measurement matrix. Secondly, the Peak Signal-to-Noise Ratio (PSNR) and reconstruction time were also compared using five reconstruction algorithms, such as the Alternating Direction Method of Multipliers (ADMM) combined with Total Variation (TV) regularization (ADMM_TV) and Subspace Pursuit (SP). Evaluation indicators show that the ADMM_TV performed best, in terms of the PSNR, Mean Squared Error (NMSE), and Structural Similarity Index (SSIM). The Natural Image Quality Evaluator (NIQE) shared the lowest value when the measurement ratio exceeded 6.0%. Although there was no outstanding advantage in the reconstruction time, the overall performance of the ADMM_TV algorithm was better than that of the rest. Finally, the Attention-Enhanced Super-Resolution Generative Adversarial Network with Multiple Scales (A-ESRGAN-Multi) model was utilized to process the reconstructed images with the different sampling rates during compressed sensing. A better performance was then achieved, compared with the Real-ESRGAN and the Attention-Enhanced Super-Resolution Generative Adversarial Network with a Single Scale (A-ESRGAN-Single). The image quality and the edge contrast were enhanced for the subsequent image segmentation. Consequently, it was very feasible to combine the compressed sensing and the A-ESRGAN-Multi model, in order to detect the fullness of sunflower seeds. The average error of plumpness on the validation set was only 2.50%, and the maximum error of detection was 6.41%. In conclusion, the combination of compressed sensing and the A-ESRGAN-Multi model can effectively save 82.5% of the sampling time and storage space. Data volume was also reduced to alleviate the pressure on the storage devices and hardware costs during data transmission and processing. Thus, the finding can also provide new technical support to the quality detection of sunflower seeds.