Detection of Pumpkin Seed Plumpness Using Terahertz Imaging Enhanced With Compressed Sensing and Super‐Resolution

The viability and market value of pumpkin seeds are critically dependent on their internal plumpness, which is a comprehensive indicator of seed quality and can be compromised by factors such as hollow kernels resulting from improper storage or processing. Traditional methods for assessing internal quality are often destructive, time-consuming, and inefficient. In this study, the internal quality of pumpkin seeds is evaluated for the first time using the nondestructive terahertz (THz) time-domain imaging system combining compressed sensing (CS) with the Real-World Enhanced Super-Resolution Generative Adversarial Network (Real-ESRGAN) method. Using the Alternating Direction Method of Multipliers-Total Variation (ADMM_TV) as the reconstruction algorithm, five measurement matrices, including the Gauss Matrix (GaussMtx) and the Part-Fourier Matrix (PartFourierMtx), were compared. After selecting GaussMtx, the performance of five reconstruction algorithms, including Basis Pursuit (BP) and Stagewise Weak Orthogonal Matching Pursuit (SWOMP), was further compared. The super-resolution reconstruction of THz images reconstructed by CS was performed through Real-ESRGAN. The image quality was evaluated by objective indicators, and the detection accuracy was verified by the plumpness error. The results indicate that the combination of GaussMtx and ADMM_TV performs best in terms of PSNR, NMSE, and SSIM. After processing with Real-ESRGAN, the edge sharpness and details of THz images were significantly improved. The fullness error was only 2.23%, and the average detection error on the verification set was 3.37%. In summary, the combination of GaussMtx and ADMM_TV reconstruction algorithm, followed by super-resolution processing with Real-ESRGAN, can effectively improve the efficiency and accuracy of pumpkin seed quality detection. PRACTICAL APPLICATIONS: This research enables seed companies and food processors to quickly and accurately identify plump, high-quality pumpkin seeds without damaging them. By using an advanced imaging technique, the method can help automate quality control on production lines, ensuring better seed selection for planting and more consistent product quality for consumers. This contributes to reducing waste and improving the overall value of the agricultural products.