基于卷积神经网络的目标检测方法

Object Detection Based on Convolutional Neural Network

摘要: 基于候选区域的目标检测算法检测精度较高，但检测速度较慢，无法达到实时检测的效果。针对这一问题，本文提出了一种新型的基于深度可分离卷积神经网络的目标检测方法。首先通过ResNet-101和深度可分离卷积层，提取目标的精简特征图，减少计算量，以提高检测速度；与此同时为了弥补提高检测速度带来的精度损失，提出采用关键点导向的策略，代替传统的回归方法，该策略利用全卷积神经网络对物体位置的敏感特性，有效保留物体的空间信息，使得算法对目标物体定位更精确。最后，为了提高算法对小目标物体的检测能力，使用PS-RoI Align方法代替传统的池化方法，在一定程度上提高算法对小目标物体的检测能力。实验结果表明，在COCO数据集上，该方法能够取得较好的检测效果。

Abstract: Object detection algorithms based on regions proposals have higher detection accuracy, but the detection speed is slower, which cannot achieve the effect of real-time detection. Aiming at this problem, this paper proposes a new type of target detection method based on deep separable convolutional neural network. First, through ResNet-101 and a deep separable convolutional layer, a simplified feature map of the target is extracted to reduce the amount of calculation to increase the detection speed; at the same time, in order to compensate for the accuracy loss caused by the increased detection speed, a key-point-oriented strategy is proposed. Instead of the traditional regression method, this strategy uses the sensitivity of the full convolutional neural network to the position of the object, effectively retains the spatial information of the object, and makes the algorithm locate the target object more accurately. Finally, in order to improve the algorithm’s ability to detect small targets, the PS-RoI Align method is used instead of the traditional pooling method to improve the algorithm’s ability to detect small targets. Experimental results show that the method can achieve better detection results on the COCO dataset.