基于图像识别的带式输送机输煤量和跑偏检测方法

传统的卷积神经网络(CNN)是单任务网络,为实现带式输送机输煤量和跑偏的同时检测,使用2个卷积神经网络分别对输煤量和跑偏进行检测,导致网络体积大、参数多、计算量大、运行时间长,严重影响检测性能.为降低网络结构的复杂性,提出了一种基于多任务卷积神经网络(MT-CNN)的带式输送机输煤量和跑偏检测方法,可使输煤量检测和跑偏检测这2个任务共享同一个网络底层结构和参数.在VGGNet模型的基础上,增大卷积核和池化核的尺度,减少全连接层通道数量,改变输出层结构,构建了MT-CNN;对采集的输送带图像进行灰度化、中值滤波和提取感兴趣区域等预处理后,获取训练数据集和测试数据集,并对MT-CNN进行训练;使用训练好的MT-CNN对输送带图像进行识别分类,实现输煤量和跑偏的准确、快速检测.实验结果表明,训练后的MT-CNN在测试数据集中检测准确率为97.3%,平均处理每张图像的时间约为23.1 ms.通过现场实际运行验证了该方法的有效性.

Detection method of coal quantity and deviation of belt conveyor based on image recognition

Traditional convolutional neural network( CNN) is a single-task network. In order to realize simultaneous detection of coal quantity and deviation of belt conveyor, two CNNs are used to detect coal quantity and deviation respectively, resulting in large network volume, many parameters, large computation and long operation time, which seriously affect detection performance. In order to reduce complexity of network structure, a detection method of coal quantity and deviation of belt conveyor based on multi-task convolutional neural network( MT-CNN) was proposed, which could make two tasks of coal quantity detection and deviation detection to share the same network underlying structure and parameters. On the basis of VGGNet model, MT-CNN is constructed by increasing scale of convolution kernel and pooling kernel, reducing the number of channels in full connection layer,and changing structure of output layer. After preprocessing the acquired conveyor belt images, such as graying,median filtering and extracting region of interest, the training dataset and test dataset are acquired, and the MTCNN is trained. The trained MT-CNN is used to identify and classify the conveyor belt images, so as to realize accurate and fast detection of coal quantity and deviation. The experimental results show that detection accuracy of the trained MT-CNN in the test dataset is 97.3%, and average processing time of each image is about 23.1 ms. The effectiveness of the method is verified by field operation.