Abstract: Earth-rock dam is one of the most widely used dam types in hydropower industry, among which rockfill dam is the best choice and its dam material particle size distribution has an important impact on construction quality and operation safety. Because of the disadvantages of manual screening, image recognition technology is becoming a hot spot. Based on image recognition, this paper carries out in-situ gradation detection of rockfill materials. According to the shape and distribution characteristics of stone particles, the rockfill materials are divided into four layers: upper, middle upper, middle lower and lower, and the four gradation curves are weighted and merged according to their respective proportions. The error between the merged curve and the screening curve is about 10%, which is reduced by 10% compared with the previous ones, and the detection results are more accurate; At the same time, an evaluation model for the particle size gradation of the dam material has been established. By comparing the fractal index and screening index of the model, the feasibility of using fractal dimension D to evaluate the gradation distribution of dam materials is verified. This method can provide reference for rapid and accurate detection of particle size distribution in similar projects.