摘要 针对使用通用目标检测方法检测稠密目标常出现的漏检的问题,提出了一种高效的基于多级多分辨网络的稠密目标检测方法。首先,通过多级多分辨的训练策略捕捉目标密集所丢失的细节。其次,利用基于形状先验的锚点生成方法,统计不同尺度下稠密目标所具有的形状变化。最后,考虑到稠密分布的目标具有较大的外观差异,通过采用不同尺寸的卷积核提取图像不同尺度的特征信息,有效解决现有检测模型中的目标信息丢失问题。在公开的车辆数据集CARPK(The Car Parking Lot Dataset)上验证了所提方法的有效性。与现有公开最好的GANet(Guided Attention Network)实验结果相比,其平均准确率提升了6.4个百分点。实验结果表明,对于不同场景条件下的稠密分布的目标,所提出的多级多分辨网络模型能够达到更好的检测效果。
Abstract:Since the dense objects were often missed when using the general detection method, an efficient dense object detection method based on a multi-level and multi-resolution network was proposed. Firstly, through the multi-level and multi-resolution training strategy, the missing details can be captured. Secondly, an effective method of anchor generation based on shape priors was used to count the shape changes of dense objects at different scales. Finally, considering that the densely distributed objects have large appearance differences, some convolution kernels of different sizes were used to extract feature information at different scales in the image, which effectively solved the problem of missing object information in the existing detection model. The effectiveness of the proposed method was evaluated on the public vehicle dataset, The Car Parking Lot Dataset (CARPK). Compared with the experimental results of the best publicly available method Guided Attention Network (GANet), Average Precision (AP) was 6.4 percentage points higher than GANet. The theoretical analysis and experimental results show that the proposed multi-level and multi-resolution network can achieve better detection results for densely distributed objects in different conditions.