With the social and economic development, the number of patients with diabetic retinopathy is increasing, and thus early diagnosis is of great significance to reduce the incidence of blindness. The hard exudate detection in the fundus is an important part of the diagnosis, but the traditional detection method is influenced by subjective factors with low accuracy and efficiency. Therefore, this study proposes a hard exudate detection algorithm based on the IHBMO-RF algorithm to assist doctors with detection. Specifically, the swarm is initialized through the introduction of the principle of the good-point set, which can not only keep the diversity of the swarm but also accelerate the convergence speed of the swarm. In this way, the problem of local optimization can be solved in machine learning. Experiments are conducted on the public fundus database DiaretDB1, and the results show that the accuracy of the proposed method reaches 95.4%. Compared with algorithms in the existing studies, the proposed algorithm has achieved a better effect, which is of certain significance for the auxiliary hard exudate detection in the fundus.