Since the 1950s, plastics have played an essential role in the world, and more than 8.3 billion tones of plastics have been produced. This number keeps increasing every year. Bio-polymers become a solution to the problem that non-degradable plastics are still used to produce frequently-used plastics products. How to sort them becomes a problem. A sorting solution neural network structure called GenNet has already developed by combine hyperspectral images with a neural network. This project is focusing on integrating a new neural network backbone to the GenNet which has already been able to learn and classify normal polymers such as PET, PEF, PP, and PE. The neural network backbone decided to integrate, called HRNet+OCR and HRNet. The HRNet+OCR achieved the highest IOU value 84.5% on semantic segmentation of public dataset Cityscapes, and it’s a derivative network from HRNet. Therefore, these two networks are decided to integrate into the GenNet and train and test on a dataset with polymers such as PET, PEF, PP, PE, and bottle made in the same material. The result is that HRNet and HRNet+OCR can not provide better performance than Deep Res U-Net and U-Net++. And the performance of HRNet+OCR will be significantly influenced by choice of the optimizer. By changing optimizer to Adam, HRNet+OCR shows a 46.128% increase on Mean IOU value on the dataset which includes plastics flakes and same material bottles. The HRNet and HRNet+OCR can learn hyperspectral data but may not be the best choice for it.