基于多维空间离散基准的目标检测网络设计

The Design of Object Detection Network Based on Multidimensional Space Discrete Datum

针对目标检测网络中检测头的设计缺少依据,不能对不同数据集进行相应优化的问题,用一至四维的空间离散基准和正样本点(组)分配冲突率模型,提出了相应的设计依据。即基于位置的分配原则中,针对特定数据集的正样本点在4个回归量所组成的空间维度中的分布特点,在不高于指定的分配冲突率的条件下,以减少正负样本不均衡度和系统计算资源为目的,设计出具有最少点数检测头的规律。其本质是一个回归基准编码和正负样本分配设计的过程,进而实现检测性能与资源消耗之间的平衡。得出基于卷积目标检测网络检测头的回归基准可以采用4个回归中的任意组合,但需结合样本分配策略与数据集的正样本分配冲突率进行综合设计的结论。

Due to lack of design basis and could not be optimized for different datasets,the design of detection heads in target detection networks was addressed by proposing a corresponding design rationale through the study of a conflict rate model for the allocation of spatially discrete datum and positive sample points(groups)in dimensions ranging from 1 to 4.Based on the principle of location-based allocation,and considering the distribution characteristics of positive sample points for a specific dataset in the spatial dimensions composed of four regression variables,a design pattern with the minimum number of points for detection heads was devised under the condition that the allocation conflict rate is not higher than the specified one.The objective is to reduce the imbalance between positive and negative samples and optimize computational resources.The process is essentially a regression benchmark encoding and positive/negative sample allocation design,aiming to achieve a balance between detection performance and resource consumption.The conclusion is that the regression benchmark based on the detection head of convolutional object detection network can use any combination of four regressions,but it needs to be comprehensively designed in combination with the sample allocation strategy and the maximum positive sample allocation conflict rate of the generated dataset.