Aluminum ingot alloy is one of the commonly used materials in industrial production and intelligent manufacturing, whose quality directly affects the performance of aluminum processed products. Therefore, the inspection of surface defects of aluminum ingot alloy is extremely valuable for actual industrial engineering. Aiming at the issues of low detecting precision and the slowly processing rate thatexisted in the traditional target detection methods for aluminum ingot alloy dataset, the YOLOv5-based improvement model RER-YOLO is proposed. Firstly, the aluminum ingot alloy dataset is coped with the image pretreatment methods of rotation, translation, contrast and brightness transformations in a random combination so as to boost the capacity of generalization for model training. Secondly, a multi-scale characteristic extraction network block (Res2Net) is utilized to take the place of the C3 block in the previous YOLOv5 to augment the model's ability that can accurately extract rich features. Finally, an over-parameterization-based re-parameterized convolutional block is utilized in place of the 3×3 convolutional blocks in the Res2Net residual block and baseline model, enlarging the search space of the network and boosting the model's fitting ability while maintaining inference rate. The comparison experimental results demonstrate that the RER-YOLO reaches a mean average precision of 75.1% on the aluminum ingot alloy dataset, which is higher 4.9% than the conventional YOLOv5 and does not increase the inference delay. It also improves the detection accuracy by 12.7% for burr defects, which are fewer in number in the dataset and the defect features are difficult to extract. It can be seen that the presented model in this study has an important reference value towards detecting surface defects in aluminum ingot alloy.