基于HSV色彩空间的图像感兴趣区域提取方法研究

Research on Image ROI Extraction Method Based on HSV Color Space

嵌入式设备中部署深度学习检测模型往往面临算力不足的问题,而感兴趣区域(ROI)提取可作为一种高效的性能优化手段。文章提出一种基于HSV(Hue,Saturation,Value)色彩空间模型的ROI提取的方法,将检测目标的像素信息转化到HSV色彩空间,在色相-饱和度(H-S)平面引入DBSCAN(Density-Based Spatial Clustering of Applications with Noise)聚类算法,精确定位目标的主色彩像素在H-S平面上的分布位置,同时过滤杂乱色彩,然后通过Quickhull(快壳)凸包算法,从散点数据中拟合出主色彩的精确分布范围。根据获取的主色彩范围对像素进行遍历,可以根据色彩信息有效地提取ROI。实验结果表明,经过该方法优化后的Faster R-CNN(Faster Regions with Convolutional Neural Networks)算法,较原模型减少了57.08%的平均推理耗时,同时精确率提升了0.9百分点。这对于嵌入式设备中进行实时目标检测具有重要的现实意义。

Deploying deep learning models for object detection on embedded devices often faces the issue of insufficient computing power,while Region of Interest(ROI)extraction can serve as an efficient performance optimization solution.This paper proposes a ROI extraction solution based on HSV(Hue,Saturation,Value)color space model,where target pixel information is transformed into HSV color space,and the DBSCAN(Density-based Spatial Clustering of Applications with Noise)clustering algorithm is introduced in the Hue-Saturation(H-S)plane to accurately locate the distribution position of the main color pixels of the target in the H-S plane,filtering out irrelevant colors.Subsequently,the Quickhull algorithm accurately defines the main colors'distribution.By traversing pixels within this range,ROI is efficiently extracted based on color information.Experimental results show that the Faster R-CNN(Faster Regions with Convolutional Neural Networks)algorithm optimized by the proposed method reduces the average inference time by 57.08%and increases the precision by 0.9 percentage points,compared to the original model.This is of great practical significance for real-time target detection on embedded devices.