The invention relates to a catenary dropper defect detection model training method and a defect detection method, and belongs to the technical field of catenary fault detection. According to the invention, a contact network dropper defect detection model based on two layers of convolutional neural networks is constructed to extract the image features of a catenary dropper, and the defects of the dropper are classified, by utilizing the advantages of a convolution algorithm, the method is not interfered by the objective conditions, such as the geometrical shapes, the shielding of the dropper, etc., in the image, and compared with a traditional detection mode using manual image watching, the robustness and the efficiency are higher. Meanwhile, on the basis of the characteristics of the dropper, the initial anchor is set by using a clustering algorithm, so that the image detection efficiency and accuracy are improved; in addition, during the defect detection, a corresponding threshold value is set for the output of each layer of model, so that the accuracy of the final dropper defect detection is improved.