基于VGGNet改进网络结构的多尺度大熊猫面部检测

大熊猫个体识别对研究大熊猫的种群数量非常重要，大熊猫面部检测是基于面部图像的大熊猫个体识别方法中的首要关键步骤。针对现有的大熊猫面部检测方法精确度不高的问题，提出基于VGGNet-16改进网络结构的多尺度大熊猫面部检测方法。首先，以VGGNet-16网络结构为基础，通过增加残差结构与BN层，降低卷积层通道数，并采用LeakyRelu激活函数等改进，构建一个新的特征提取主干网络。其次，将一个3尺度的特征金字塔网络结构与SPP结构结合用于目标检测。最后，使用深度分离卷积结构替代常规卷积结构。实验结果表明，提出的大熊猫面部检测方法在测试集上能够达到99.48%的mAP，检测性能优于YOLOv4。

Multi-scale giant panda face detection based on the improved VGGNet architecture

Individual identification of giant pandas is very important for studying their population of them.. Giant panda face detection is the first key step of giant panda individual identification method based on facial images. To solve the problem that the precision of the existing giant panda face detection methods are low, a multi-scale giant panda face detection method based on improved VGGNet-16 architecture was proposed in this paper. Firstly, based on the VGGNet-16 network architecture, a new feature extraction backbone network was constructed through certain improvements such as adding the residual block and BN(Batch Normalization) layer, reducing the channel dimensionality of convolution layer and adopting LeakyRelu active function as well. Secondly, a 3-scale feature pyramid network structure was combined with SPP(Spatial Pyramid Pooling) structure for object detection. Finally, the conventional convolution architecture was replaced with the depwise separation convolution architecture. Experimental results show that the proposed method can achieve 99.48% mAP(mean average recision) in the test dataset, and the detection performance is better than YOLOv4(You Only Look Once Version 4).