Scholarlink

There are myriad varieties of rice grains in different parts of the world grown on diverse climates, physiographies and terrains. As sorting of these varieties by physical means is cumbersome, innovative image processing techniques using AI, ML and DL are currently providing newer insights and accurate distinguishing mechanisms. In this paper, the proposed model transforms a low-resolution rice image to an enhanced super resolution image by Deep Learning. The model is christened as a ‘Twenty-Fold Protracted Residual Network’ (TFPRN) to indicate the twenty-stage Protracted Residual Module (PRM)s are employed here. The model is trained on 80% of the total 75,000 raw images of rice, with 5 classes each of 15,000 low resolution images so to test them randomly on the remaining 10% of the images and validate on 10% of the residual set of images. MSE, PSNR and SSIM are computed to quantify the loss function, to train the model and to validate the model, respectively. The model did gain a good accuracy level of 99.14, 98.45 and 97.82% and PSNR of 34.468, 32.137 and 31.467 and also SSIM of 0.9124, 0.8938 and 0.8683, respectively, on Super Resolution images upped by factors of 2, 4 and 6.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Twenty-Fold Protracted Residual Network (TFPRN) for Rice Grain Type Detection

  • 章节

  • 发表日期：

    2026年

查看全文
  • P. V. Yeswanth,
  • S. Deivalakshmi

摘要

There are myriad varieties of rice grains in different parts of the world grown on diverse climates, physiographies and terrains. As sorting of these varieties by physical means is cumbersome, innovative image processing techniques using AI, ML and DL are currently providing newer insights and accurate distinguishing mechanisms. In this paper, the proposed model transforms a low-resolution rice image to an enhanced super resolution image by Deep Learning. The model is christened as a ‘Twenty-Fold Protracted Residual Network’ (TFPRN) to indicate the twenty-stage Protracted Residual Module (PRM)s are employed here. The model is trained on 80% of the total 75,000 raw images of rice, with 5 classes each of 15,000 low resolution images so to test them randomly on the remaining 10% of the images and validate on 10% of the residual set of images. MSE, PSNR and SSIM are computed to quantify the loss function, to train the model and to validate the model, respectively. The model did gain a good accuracy level of 99.14, 98.45 and 97.82% and PSNR of 34.468, 32.137 and 31.467 and also SSIM of 0.9124, 0.8938 and 0.8683, respectively, on Super Resolution images upped by factors of 2, 4 and 6.