为了解决浮选气泡图像现场光照不均、相互黏结、无背景等造成的直接形态特征提取困难问题, 提出一种浮选气泡NSCT域(Nonsubsampled Contourlet Transform, NSCT)的多尺度等效形态特征提取及识别方法。通过NSCT变换将浮选气泡图像分解为低频图像和多尺度多方向高频图像; 采用模糊集方法二值化低频子带图像, 得到气泡亮点图像, 提取亮点个数、平均面积、标准差和椭圆率作为等效形态尺寸特征; 结合方向模极大值及差分盒维法计算各高频子带方向的分形维数; 最后, 将多尺度多方向等效形态尺寸特征作为输入, 采用量子门节点神经网络对三类浮选气泡图像进行状态识别和分类。实验结果表明, 该方法提取的等效形态尺寸特征与分类的相关性高, 能对三种类型浮选气泡图像进行有效的状态识别, 平均识别准确率达95.1%。本算法的识别准确率较几种流行算法而言有较大提高, 适用于动态变化的浮选工况。

To solve the problem of cohesion and background-free and uneven illumination, which makes it difficult to extract direct morphological features from flotation bubble images, a multi-scale equivalent morphological feature extraction and recognition method for flotation bubbles was proposed in a nonsubsampled contourlet transform (NSCT) domain. Firstly, the flotation bubble image was decomposed via NSCT to obtain a low frequency subband and multi-scale and multi-directional high frequency subbands. The fuzzy set method was used for the binarization of the low frequency subband image to obtain the bubble bright spot image, and the number of bright spots, average area, standard deviation, and ellipticity were extracted as the equivalent morphological size features. Thereafter, the directional modulus maxima and differential box-counting method were used to calculate the fractal dimensions of the high frequency subband directions. Finally, by using the multi-scale and multi-directional equivalent morphological size features as the input, the state recognition and classification of three types of flotation bubble images were carried out via a quantum gate node neural network. The experimental results show that the extracted equivalent morphological size features are highly correlated with the classification and it can be effectively used to recognize the state of three types of flotation bubble images. The average recognition accuracy is 95.1%, which is higher than that of several common algorithms, and it is suitable for dynamic flotation conditions.