Detection the internal quality of watermelon seeds based on terahertz imaging combined with image compressed sensing and improved-real-ESRGAN

Abstract

Traditional watermelon seed detection methods are inefficient, destructive and difficult to accurately identify empty shells or defective seeds. It limits the improvement of market value and planting efficiency. In this study, a non-destructive detection method based on terahertz (THz) imaging combined with compressed sensing and Improved-Real-ESRGAN is proposed for fast and accurate assessment of the internal quality of watermelon seeds. The study is divided into three stages: firstly, THz imaging efficiency is optimised by combining compressed sensing techniques. The THz time-domain imaging system is utilised to acquire images of the interior of the seeds. Different random measurement matrices with sparse reconstruction algorithms are compared. The imaging efficiency and reconstruction quality are optimised. Second, the image quality is improved by super-resolution enhancement modelling. The discriminator upsampling method in Real-ESRGAN is replaced with pixel shuffling. Super-resolution enhancement model is built. This step provides a clearer image input for subsequent seed fullness detection. Thirdly, threshold segmentation is used to calculate seed fullness based on the enhanced image. Its applicability in defect detection is verified. The results show that the combination of BernoulliMtx and ADMM_TV algorithms saves 87.5 % of the imaging time with a measurement ratio of only 12.5 %. THz image after compressed perception combined with Improved-Real-ESRGAN processing: PSNR, SSIM and other indicators show that the image quality have been significantly improved, as well as the authenticity of image details. The fullness detection error is reduced to 3.52 %, and the average detection error of the validation set is 5.79 %. Efficient identification of seed defects is achieved. Compared to the traditional method, the optimised process proposed in this study significantly reduces the detection time. The detection accuracy is improved. It provides theoretical support and technical reference for the quality detection of agricultural products, especially for the evaluation of internal defects of seeds.