Abstract: In the process of converter steelmaking,there is a close relationship between the color and texture of converter mouth flame and the carbon content of molten steel at the endpoint .To extract effective flame image features for accurately predicting the carbon content of the molten steel,an Improved Multi-trend Binary Coded Descriptor method (IMTBCD) was proposed by combining multi-scale with multi-direction characteristics of the flame textures.Through the color channel fusion strategy with the correlation between color channels,the color texture representation of the flame image was obtained.The multi-scale non-uniform sampling strategy was used to select sampling points in each scale range to construct the multi-scale expression of color textures.According to the different changing of the sampling points from the center point symmetry direction and the diagonal symmetry direction,the color texture features were obtained through multi-trend coding.With the extracted color texture features,the carbon content was predicted via Generalized Regression Neural Network (GRNN) regression model,and the experiment showed that the accuracy of carbon content prediction was 95.7% within the error range of 0.02%.

Key words: flame images, carbon content prediction, multi-scale non-uniform sampling, pixels change trend, color texture features

摘要： 转炉炼钢终点时刻炉口火焰的颜色和纹理与钢水碳含量之间存在对应关系,为了提取有效的火焰图像特征以准确预测钢水碳含量,结合火焰纹理多方向多尺度等特点,提出一种改进的多趋势二进制编码彩色纹理特征表述方法。考虑颜色通道间的相关性,通过颜色通道融合策略得到火焰图像的彩色纹理表示;采用多尺度非均匀采样策略选取各尺度范围内的采样点来构建彩色纹理的多尺度表达;根据中心点对称方向和对角线对称方向上采样点不同的变化进行多趋势编码,得到彩色纹理特征,选用广义回归神经网络模型预测碳含量。实验表明,碳含量预测在误差范围0.02%以内的准确率为95.7%。

关键词: 火焰图像, 碳含量预测, 多尺度非均匀采样, 像素变化趋势, 彩色纹理特征