基于Mask RCNN的滤袋开口检测方法

A detection method of filter bag opening based on Mask RCNN

摘要:

滤袋开口的检测与定位在滤袋智能生产过程中占据着至关重要的地位。但由于滤袋具有柔性的特点,常规检测方法很难有效进行,且定位精度也不能满足生产要求。本文提出一种结合可变形卷积与掩码信息的多尺度目标检测器,该检测器使用可变形卷积改进主干网络高层中的固定卷积,结合特征金字塔技术实现多尺度信息融合。然后将所得多尺度信息通过区域提议网络生成候选区域,采用改进的Soft-NMS方法进行筛选,最终送入检测头进行识别与分割。本文在滤袋图像数据集上进行了实验。结果表明,提出的算法实现了滤袋开口的准确识别与高精度定位。

Abstract:

The detection and positioning of the filter bag opening occupies a crucial position in the intelligent production process of the filter bag. However, due to the flexible of the filter bag, it is difficult to carry out conventional detection methods effectively, and the positioning accuracy cannot meet the production requirements. In this paper, a multi-scale target detector combining deformable convolution and mask information is proposed. This detector uses deformable convolution to improve fixed convolution in the high layer of the backbone network, combined with feature pyramid technology to achieve multi-scale information fusion. The obtained multi-scale information is then used to generate candidate regions through the region proposal network, which is screened using the improved Soft-NMS method, and finally sent to the detection head for recognition and segmentation. In this paper, experiments were carried out on the filter bag image data set. The results show that the proposed algorithm achieves accurate identification and high-precision positioning of the filter bag opening.