基于VDNet卷积神经网络的羊群计数

为了避免传统羊群计数任务中，羊只之间相互遮挡带来的干扰，提高羊群计数的准确度，采用了视觉几何群(VGG-16)与空洞卷积(DC)相结合的VDNet神经网络羊群计数方法。该方法在网络前端采用去除了全连接层的VGG-16网络提取2-D特征，后端采用6层具有不同空洞率的DC提取更多的高级特征；DC在保持分辨率不变的同时扩大了感受野，替代池化操作，降低了网络的复杂性；最后用一层卷积核大小为1×1的卷积层输出高质量的密度图，通过对密度图像素积分得出输入图片中羊的数量，并进行了理论分析和实验验证。结果表明，VDNet的平均绝对误差为2.51，均方误差为3.74，平均准确率为93%。这一结果对羊群计数任务是有帮助的。

Herd counting based on VDNet convolutional neural network

In order to avoid the interference of mutual occlusion between sheep in the traditional flock counting task and improve the accuracy of flock counting, the VDNet(VGG-16+DC net) convolutional neural network flock counting method, combining visual geometry group(VGG) 16 and dialated convolution (DC) net, was adopted. VGG-16 with the fully connected layer removed was used at the front end of the network to extract 2-D features, 6 layers of DC with different dilated rates was used to extract more advanced features. DC expanded the receptive field, replaced the pooling operation, and decreased the complexity of the network while kept the resolution unchanged at the same time. The theoretical analysis and experimental verification were carried out. Finally, a convolutional layer with a convolution kernel size of 1×1 was used to output a high-quality density map, and then the number of sheep in the input image was obtained by integrating the pixels of the density map. The results show that the average absolute error of the counting method in this paper is 2.51, the mean square error is 3.74, and the average accuracy is 93%, respectively. This result is helpful for the task of counting sheep.