For the difficulties in identifying and locating surface defects on mining conveyor belt, a mining Conveyor Belt Defect Detection System is designed based on maximizing reliable region through decision-level fusion. Firstly, the YOLOv4 and SSD models are trained using public datasets with the transfer learning, and the decision-level fusion method is proposed for defect detection results based on its predicted score and overlapping situation. At the decision level, the two models are fused to maximize reliable regions, which can achieve defect identification and localization of conveyor belt. The effectiveness and reliability of the proposed method are validated and analyzed through public datasets and mining conveyor belt defect datasets. The results illustrated that the proposed method can fully utilize the detection results of YOLOv4 and SSD models to achieve high accuracy, recall, and overlap rates. It can achieve the accuracy rate and overlap rate of over 85% and 0.6, respectively. The defect detection system is beneficial for defect detection and maintenance of mining conveyor belt, and it is significate for safe and reliable operation of the mining conveyor.