多层卷积网络目标识别算法在烟丝宽度检测中的应用研究

Application of multi-layer convolutional networkobject recognition algorithm in tobacco width detection

为了提高烟丝宽度检测效率,提出一种多层卷积网络目标识别算法。该算法以YOLOv3为基础结构,通过添加特征层、注意力机制、空间金字塔结构、改进聚类函数和损失函数的方式,实现了对传统识别算法的改进优化。基于多层卷积网络目标识别算法构建烟丝宽度检测模型,通过获取图像中目标烟丝宽度,完成对烟丝宽度的精准检测。烟丝的主客观识别效果与宽度检测结果表明:该算法能够较好地保留烟丝轮廓细节,具有较强的识别能力和较高的检测精度,交并比、精准度、AUC分别达到了0.966、0.969、0.963,且检测最大误差为-0.073 mm,相对误差较低。实验结果证实该算法对烟丝宽度检测具有较高的实用价值,大大提高了检测精度。

In order to improve the efficiency of tobacco width detection,a multi-layer convolutional network target recognition algorithm is proposed.This algorithm is based on YOLOv3 and applies the methods of adding feature layers,attention mechanism,spatial pyramid structure.The system improves the clustering function and loss function to optimize traditional recognition algorithms.A tobacco width detection model is constructed based on a multi-layer convolutional network target recognition algorithm.By obtaining the target tobacco width in the image,accurate detection of tobacco width is achieved.The subjective and objective recognition effects of tobacco and the width detection results show that the algorithm can well preserve the details of tobacco contour,have strong recognition ability and high detection accuracy intersection to union ratio,the accuracy AUC reaches 0.966,0.969 and 0.963 respectively with a maximum detection error of-0.073 mm.The experimental results confirm that the algorithm has high practical value for tobacco width detection,greatly improving detection accuracy.