Abstract: Under the background of China's full efforts to promote green and low-carbon metallurgy, the steel industry is facing transformation challenges and green development requirements. This urgently requires in-depth adjustment and optimization of the blast furnace burden structure to achieve the dual goals of energy conservation, carbon reduction, and pollution emission reduction. In modern metallurgical production, compared with sinter, pellet has better energy conservation and emission reduction in its own production. Accurate detection of the quality and quantity of pellets is crucial. In steel manufacturing enterprises, operators rely on empirical means such as screening detection to evaluate the particle size state of pellets, which has a long measurement time and low accuracy. It is particularly urgent and important to adopt a fully automated monitoring and measurement system for optimized management of pellet particle size. An algorithm combining YOLOv5+U-Net is proposed. By comparing with the neural network models of Deeplabv3+ and FCN, the advantages and limitations of each of the three algorithms are obtained. The Deeplabv3+ algorithm adopts a convolution-free kernel and a convolution-free-based spatial pyramid pooling （ASPP） architecture. The dilated convolutions with larger and smaller sampling rates can capture the whole and detailed parts of pellets. The FCN algorithm has certain robustness to noise, deformation, and changes in ambient light in images. It overcomes interference factors through learned features and accurately performs segmentation. After years of iteration and development, the YOLO algorithm transforms object detection into a regression problem and introduces normalization, further improving the detection accuracy. A neural network model based on the YOLOv5+U-Net algorithm is adopted by this paper and its use of computer deep learning and image processing technology to detect and identify pellet particle sizes is verified. In the test of detecting the particle size of green pellets, the accuracy rate reaches more than 97% and the recall rate reaches 66.8%. The pellet detection method proposed in this paper has broad application space in fields such as industrial automation and quality control, providing new ideas and methods for the development of pellet detection technology.