基于像素偏转模型和机器学习的室外图像天空像素检测

Pixel detection in outdoor image sky based on pixel deflection model and machine learning

图像中的天空区域对于基于视觉的地面机器人导航具有重要意义,为了识别图像中的天空部分,本文提出了一种基于像素偏转模型的BP神经网络天空识别方法。首先,制作天空图像集和非天空图像集,天空图像集由各种天气情况下的天空提取而成,非天空图像集由非天空的景物构成,主要包括建筑、汽车、树木、植物等;其次,使用提出的像素偏转模型提取天空图像集和非天空图像集的像素特征并进行处理,对天空像素点和非天空像素点进行标注,利用BP神经网络对像素特征进行训练,得到权重文件;最后,使用得到的权重文件进行天空的识别。为了更好的说明本文算法和模型的优越性,使用本文算法与Otsu算法、Ye Hu算法、Graph-cut算法和Mask-Rcnn算法模型进行了比较,并设计了两组组对比实验,第一组实验进行识别效果的主观评价,第二组实验利用Cam Vid数据集的天空类进行算法精度的定量分析。

The sky region in the image is of great significance for vision-based ground robot navigation.In order to identify the sky part in the image,this paper proposes a BP neural network sky recognition method based on pixel deflection model.Firstly,sky image sets and non-sky image sets were made.Sky image sets were extracted from the sky under various weather conditions.Non-sky image sets were composed of non-sky objects,including buildings,cars,trees,plants,etc.Then,the proposed pixel deflection model is used to extract and process the pixel features of sky image set and non-sky image set,and the sky pixel points and non-sky pixel points are labeled.Then,the BP neural network is used to train the pixel features and obtain the weight file.Finally,the weight file is used to identify the sky.In order to better illustrate the advantages of the algorithm and model in this paper,the algorithm in this paper was compared with Otsu algorithm,Ye Hu algorithm,Graph-cut algorithm and Mask-RCNN algorithm.In addition,two groups of comparative experiments were designed.The first group of experiments was used for subjective evaluation of recognition effect,and the second group of experiments was used for quantitative analysis of algorithm accuracy using sky class of Cam Vid data set.