基于环形特征的卷积神经网络轮毂识别

Wheel Hub Identification of Convolutional Neural Networks Based on Ring Features

针对实际生产中不同种类轮毂的混流生产问题,提出了一种基于环形特征的卷积神经网络轮毂识别算法。将直角坐标下的环形轮毂映射到极坐标中,归一化为标准形式的矩形,提取轮毂图像的环形特征信息,减少冗余特征产生的影响;设计了一种改进的VGG网络架构,利用深度可分离卷积打破输出通道维度与卷积核大小的联系,在不损失网络性能的同时降低了计算量,能够在实际生产中轮毂识别任务在有限的算力情况下实时进行计算;从有效性和实时性两个方面对轮毂识别算法进行评估,且通过Inception V3、SVM、KNN等模型的对比实验,验证了该算法可以实时地对轮毂自适应分类。实验表明:该方法对轮毂图像的处理精度达到99%以上,单幅图像平均处理时间降低至11.78 ms。

Aiming at the mixed-flow production of different types of wheels,a wheel hub identification algorithm of convolutional neural networks based on ring features is proposed.The circular hub in rectangular coordinates is mapped to polar coordinates,normalized to a rectangle in standard form,and the feature information is extracted by rotation to reduce the influence of redundant features.An improved VGG network architecture is designed,which uses depthwise separable convolution to break the relationship between the output channel dimension and the size of the convolution kernel.It reduces the computation without losing the network performance.The hub recognition algorithm is evaluated in terms of effectiveness and real-time performance,and through comparative experiments of models such as Inception V3,SVM,and KNN etc.The experiment shows that the method has a processing accuracy of more than 99%,and the processing time of a single image is reduced to 11.78 ms.