吴庆岗, 张卫国, 赵进超, 张秋闻, 景雨. 基于随机森林和多特征融合的青苹果图像分割[J]. 信阳师范学院学报(自然科学版), 2018, 31(4): 681-686. DOI: 10.3969/j.issn.1003-0972.2018.04.032
引用本文: 吴庆岗, 张卫国, 赵进超, 张秋闻, 景雨. 基于随机森林和多特征融合的青苹果图像分割[J]. 信阳师范学院学报(自然科学版), 2018, 31(4): 681-686. DOI: 10.3969/j.issn.1003-0972.2018.04.032
WU Qinggang, ZHANG Weiguo, ZHAO Jinchao, ZHANG Qiuwen, JING Yu. Green Apple Image Segmentation Based on Multi-feature Fusion and Random Forest[J]. Journal of Xinyang Normal University (Natural Science Edition), 2018, 31(4): 681-686. DOI: 10.3969/j.issn.1003-0972.2018.04.032
Citation: WU Qinggang, ZHANG Weiguo, ZHAO Jinchao, ZHANG Qiuwen, JING Yu. Green Apple Image Segmentation Based on Multi-feature Fusion and Random Forest[J]. Journal of Xinyang Normal University (Natural Science Edition), 2018, 31(4): 681-686. DOI: 10.3969/j.issn.1003-0972.2018.04.032

基于随机森林和多特征融合的青苹果图像分割

Green Apple Image Segmentation Based on Multi-feature Fusion and Random Forest

  • 摘要: 针对自然环境下青苹果图像中目标与背景颜色差异小和分割难度大的问题,提出一种基于多特征融合的随机森林(Random Forest,RF)分割方法.首先,基于灰度共生矩阵提取青苹果图像的能量、熵、对比度、相关性、熵标准差和对比度标准差6个纹理特征;然后,针对同幅青苹果图像提取RGB空间中的G+0.5R-B分量和HSI空间中的S+I分量作为组合颜色特征,以规避天空和高光区域对分割结果的影响;接着,以像素为单位对提取的纹理特征和颜色特征进行融合;最后,在融合特征的基础上采用随机森林对青苹果图像进行分割,并与传统仅利用单一特征的分割算法进行对比.实验结果表明,基于多特征融合的随机森林算法比传统仅利用纹理特征的算法正确分割率要高22.18%.

     

    Abstract: In natural green apple images, the color difference between targets and background is small, which increases the difficulty of segmenting out interested objects. To overcome this problem, a novel apple image segmentation method based on multi-feature fusion and Random Forest (RF) is proposed. Firstly, six texture features of energy, entropy, contrast, correlation, entropy, the standard deviations of entropy and contrast are extracted from the raw image based on Gray Level Co-occurrence Matrix (GLCM). Secondly, the color components of G+0.5R-B in RGB space and S+I in HSI space are computed as the combined color features to avoid the influence of sky and highlight area to the segmentation result. Then, the extracted texture features and color features are fused together in pixels. Finally, the image segmentation is performed by applying Random Forests to the fused features and the segmentation results are compared with the traditional method which only considers the unique kind of features. Extensive experimental results qualitatively and quantitatively demonstrate the advantages of our algorithm by a significant improvement of 22.18% segmentation accuracy than the one only with texture features.

     

/

返回文章
返回