基于可变空间感知的目标检测算法

Object detection algorithm based on variable spatial perception

检测场景中目标数量不定且随机分布,目标之间尺寸差异大,使得检测难度增加。为此,文中利用可变形卷积对位置偏差的学习能力,构建两个能够建模局部几何特征的可变空间感知模块(Variable Spatial Perception Module, VSPM)。VSPM1用于特征下采样阶段,减少分辨率降低引起的信息损失,从而有效提升检测器的颈部特征融合能力,使输入检测头的特征包含更多有益预测的信息。VSPM2用于检测头部分,通过大核卷积获取全局信息,另外,通过解耦的检测头解决分类和回归任务之间的冲突。所提算法在PASCAL VOC数据集上检测精度达到84%,相比基准算法YOLOv4提高2%,能够有效提高检测性能。

Due to the variable number and random distribution of objects in the detection scene and the significant size difference between objects,the difficulty of detection is increased.Two variable spatial perception modules(VSPM)that can model local geometric features are constructed by means of the learning ability of deformable convolution to position deviation.VSPM1 is used in the feature downsampling stage to reduce the information loss caused by the resolution reduction,so as to effectively improve neck feature fusion ability of the detector and make the features inputted into detection head contain more useful predictive information.VSPM2 is used for the detection head part to obtain global information by means of large kernel convolution,and the conflicts between classification and regression tasks are resolved by means of decoupled detection heads.The accuracy rate of the proposed algorithm on the PASCAL VOC dataset can reach 84%,which is increased by 2%compared with the baseline algorithm YOLOv4,indicating that the proposed algorithm can effectively improve detection performance.