Coal-gangue image recognition in fully-mechanized caving face based on random forest
-
摘要: 针对目前综放工作面煤矸图像识别方法存在的参数调节难度高、预测准确率低、易过拟合等问题,提出了一种基于随机森林(RF)算法的综放工作面煤矸图像识别方法。以担水沟煤矿6203综放工作面为工程背景,采集放煤口的煤矸图像并对其进行裁剪、灰度转化、对比度增强、图像滤波预处理;采用灰度-梯度共生矩阵提取出15个煤矸图像纹理特征;采用RF算法对15个煤矸纹理特征的重要性进行排序,并选取前5个实现降维处理,分析降维前后RF算法对煤矸图像的识别效果。结果表明,在决策树个数为150、采用logM2+1方法计算每次分裂时的特征数情况下,降维后RF模型的煤矸分类准确率为97%,比降维前提高4%,煤矸分类查准率为0.98,查全率为0.96,且袋外错误经50次迭代达到9%,泛化能力更强。Abstract: Aiming at problems of high difficulty in parameter adjustment, low prediction accuracy and easy over-fitting in present coal-gangue image recognition methods in fully-mechanized caving face, a coal-gangue image recognition method in fully mechanized caving face based on random forest (RF) algorithm is proposed. Taking 6203 fully-mechanized caving face of Danshuigou Coal Mine as project background, coal-gangue image of caving mouth are collected and pre-processed by clipping, gray conversion, contrast enhancement and image filtering. Fifteen texture features of coal-gangue image are extracted by gray-gradient co-occurrence matrix. RF algorithm is used to rank the importance of the fifteen coal-gangue texture features, and the first five features are selected for dimension reduction. Recognition effect of RF algorithm on coal-gangue images before and after dimension reduction is analyzed. The results show that when the number of decision tree is 150 and the number of features in each split is calculated by logM2+1 method, accuracy rate of coal-gangue classification of RF model after dimension reduction is 97%, which is 4% higher than that before dimension reduction, accuracy rate coal-gangue classification is 0.98, recall rate is 0.96, and out-of-bag error rate reaches 9% after 50 iterations with stronger generalization.
点击查看大图
计量
- 文章访问数: 111
- HTML全文浏览量: 16
- PDF下载量: 15
- 被引次数: 0