随机缩放混合与跨尺度特征增强的任务对齐目标检测算法

Task-aligned Object Detection Algorithm Based on Random Scaling Mixture and Cross-Scale Feature Enhancement

摘要:

针对TOOD算法鲁棒性差、特征金字塔顶层丢失部分语义信息、不同尺度特征层存在语义差距的问题,提出随机缩放混合与跨尺度特征增强的任务对齐目标检测算法；该算法提出2×4混合增强方法,丰富训练样本,提高模型的泛化性和鲁棒性；构造多重残差特征增强模块,自适应融合顶层不同尺度的上下文信息,减少最高层语义信息的损失；构建堆叠金字塔卷积模块,缩小不同尺度特征层之间的语义差距,提升多尺度特征的融合效果；Pascal VOC数据集上的实验结果表明,所提算法的均值平均精度、查准率、查全率分别比TOOD算法提高了3.76%、15.71%、6.28%；而且该算法的F1值与均值平均精度均优于6种主流对比算法。

Abstract:

A novel algorithm named as task-aligned object detection algorithm based on random scaling mixture and cross-scale feature enhancement is proposed to avoid the disadvantages of poor robustness, loss of some semantic information in the top-level feature layer of the feature pyramid network, and existing the semantic gap for the feature layers with different scales in the TOOD algorithm. After utilizing the 2x4 hybrid augmentation method to enrich the training samples, the presented method improves generalization and robustness of the model. In order to reduce the loss of the semantic information at the highest level, the multiple residual feature enhancement module is constructed by adaptively fusing the context information with different scales at the top level. Moreover, the proposed method constructs the stacked pyramid convolution module, which reduces the semantic gap among the different scale features and improves the effect of multi-scale feature fusion. The experimental results show that the mean average precision, the precision and the recall of the proposed algorithm on the Pascal VOC dataset are 3.76 %, 15.71 % and 6.28 % respectively higher than the TOOD algorithm. Compared with the state-of-the-art algorithms, the presented algorithm achieves higher F1 value and mean average precision.