Abstract: Aiming at the problems of tiny defects in complex castings,complex background interference,and difficulty in high-precision defect detection,a casting defect detection model named single-shot Refinement neural network for Aluminum Casting Tiny Defects Detection (Refine-ACTDD) based on deep learning was proposed.Based on the single-shot Refinement neural network for Object Detection (RefineDet) algorithm,a close-packed anchor design method was adopted to improve the accuracy of small defects,and an attention mechanism was introduced to reduce the interference of complex background detection.At the same time,a method that combined deep learning and contour discovery was proposed to realize end-to-end detection of small sample defects.After data collection,a large data set ALU-DEF containing 7816 images of appearance defects of castings was produced.The comparative training method was used on the data set to train and test.Experimental results showed that the algorithm could achieve an average accuracy of 95.44%,which had a higher Mean of Average Precision (MAP) and accuracy than faster-RCNN and YOLOv3.

Key words: deep learning, defect detection, aluminum alloy die castings, tiny defects

摘要： 针对复杂铸件外观缺陷体积小、背景复杂干扰大,较难实现高精度缺陷检测的问题,提出一种基于深度学习的铸件外观缺陷检测模型(Refine-ACTDD)。该模型基于RefineDet算法,采用密排的锚点设计方法提高微小缺陷的正检率,并引入注意力机制减少复杂背景对检测的干扰。同时,提出一种将深度学习与轮廓发现相结合的方法实现对小样本缺陷的端到端检测。经过数据采集,制作包含7 816张铸件外观缺陷图片的大型数据集ALU-DEF。最后,采用对比训练方法在数据集上进行训练和测试。实验结果表明,该算法能够达到95.44%的平均正检率,相比于Faster-RCNN和YOLOv3算法具有更高的平均精确率(MAP)和正检率。

关键词: 深度学习, 缺陷检测, 铝合金压铸件, 微小缺陷