煤矸石识别技术作为煤炭产业向智能化、绿色化转型的核心技术之一，能够提高产煤量，节约运输成本，提高煤炭生产效益。为了使煤矸石识别方法适用不同的复杂环境，必须研究增强煤矸石识别健壮性的方法。本文主要从图像预处理、特征选择与降维和煤矸石识别三个方面对增强煤矸识别的健壮性展开研究。

针对煤和矸石图像对比度较弱、光线分布不均的情况，首先采用直方图均衡技术对图像进行增强，提高了图像的对比度，为特征提取和煤矸石识别奠定良好的基础。其次，为了更加深入地提取不同光照、蒙灰以及不同背景情况下的煤矸石图像特征，除了选择基于灰度直方图的灰度特征和基于灰度共生矩阵的纹理特征，还增加了局部二值模式和方向梯度直方图两类特征。

针对图像特征维度过大的问题，提出一种基于二进制狮群算法的特征选择与降维方法。该方法对狮群算法借助S型传递函数解决特征选择和降维。实验结果表明，能将三组不同场景的煤矸石图像各自产生的20376维特征减少至40维、3118维和24维，且在使用支持向量机作为分类器的情况下，将识别准确率分别提升至88.79%、82.73%和86.96%，相较于Relief等算法表现出稳定的性能。

提出一种改进的麻雀搜索算法，主要采用改进型Sine混沌映射，可以有效地消除初始种群分布不均和遍历性较低的问题，并且将自适应动态权重因子融入其中，从而有效地缓解全局和局部搜索之间的失衡。然后将改进的麻雀搜索算法应用到支持向量机惩罚因子C和核函数参数sigma组合寻优中，提高了支持向量机对煤矸石识别的健壮性。通过与四种煤矸石识别方法在三组不同数据集的对比实验表明本文提出的方法相对于其它算法拥有更强的健壮性，三组数据集上的识别率分别为98.11%、97.27%和99.13%。

综上所述，本文从图像预处理、特征选择与降维和识别方法三个部分增强了煤矸石识别的健壮性，效果显著。

Coal and gangue identification technology, as one of the core technologies for the  intelligent and green coal industry, can improve coal production, save transportation cost and increase coal production efficiency. In order to make the coal gangue identification method suitable for different complex environments, it is necessary to study methods to enhance the robustness of coal gangue identification. This thesis mainly conducts research on enhancing the robustness of coal gangue recognition from three aspects: image preprocessing, feature selection, and dimensionality reduction.

 In view of the weak contrast and uneven light distribution of coal and gangue images, firstly, histogram equalization technique is used to enhance the images, which improves the contrast of images and lays a good foundation for feature extraction and gangue recognition. Secondly, in order to further extract the features of coal gangue images under different lighting, graying and different backgrounds, in addition to selecting the gray-scale features based on the gray-scale histogram and the texture features based on the gray-scale co-occurrence matrix, two types of features, namely, Local binary patterns and directional gradient histogram, are also added.

A feature selection and dimensionality reduction method based on binary lion swarm algorithm is proposed to address the problem of excessive dimensionality of image features. This method achieves dimensionality reduction by selecting the features searched by the lion swarm algorithm based on the decision of the S-type transfer function. The experimental results show that it can reduce the 20,376-dimensional features generated by each of three different sets of coal gangue images to 40, 3,118 and 24 dimensions, and improve the recognition accuracy to 88.79%, 82.73% and 86.96%, respectively when using support vector machines as classifiers. Compared with algorithms such as Relief, it shows stable performance.

An improved sparrow search algorithm is proposed, which mainly uses the improved Sine chaotic map, which can effectively eliminate the problems of uneven initial population distribution and low ergodicity, and incorporate the adaptive dynamic weight factor into it, thus effectively alleviating the imbalance between global and local search. Then, the improved sparrow search algorithm was applied to the combination optimization of support vector machine penalty factor C and kernel function parameter sigma, improving the robustness of support vector machine for coal gangue recognition. The comparative experiments with four coal gangue recognition methods on three different datasets show that the proposed method has stronger robustness compared to other algorithms. The recognition rates on the three datasets are 98.11%, 97.27%, and 99.13%, respectively.

In summary, this thesis enhances the robustness of coal gangue recognition with significant results from three aspects: image preprocessing, feature selection and dimensionality reduction, and recognition methods.