煤矸分选是煤矿开采过程中的重要环节，煤矸石中的化学成分混入煤炭会降低其燃烧效率和利用率，并在燃烧过程中产出大量危害人体和污染环境的有害气体。传统煤矸识别方法存在成本过高、操作困难等问题，难以满足煤矸分选的实际需求。随着深度学习技术在各个领域广泛应用，利用神经网络能够更精确地提取图像中的丰富细节信息，从而显著提升识别效率。因此，基于深度学习的煤矸石识别成为解决煤矸识别的重要手段，然而还存在模型识别精度不高、多目标识别不理想等问题。本文基于图像分类模型和语义分割模型，对煤矸识别中的单目标分类和多目标分割方法进行研究。主要研究工作有：

（1）针对实际生产环境中煤与煤矸石的多尺度形态，已有的选煤方法在煤矸识别上效率低等问题，结合图像数据预处理和图像分类模型，提出了一种基于改进EfficientNetV2的煤与煤矸石单目标识别模型。首先，选择EfficientNetV2为基础模型，在模型训练前，先使用基于管道的数据增强方法和图像处理技术扩充数据集，以模拟实际工况下煤与煤矸石的生产环境。其次，改进原始模型中的MBConv和Fused-MBConv模块，引入CAM注意力机制模块以及Hardswish激活函数，解决原有注意力机制模块与激活函数的缺陷，同时提升了网络识别精度。最后，通过优化网络结构超参数，在保持模型精度的同时加快了网络的训练速度，将改进后的模型应用于经过数据预处理之后的煤与煤矸石数据集中。实验结果表明，改进后的EfficientNetV2煤矸识别模型易训练、训练速度快、收敛速度快且识别精度高，与原始模型相比，识别准确率提高了3.98%，达到了98.24%。此外，训练速度也有所提高，改进模型的推理时间减少了6.6毫秒，由原先的每秒推理39张图片增加到了每秒推理52张图片，验证了改进后模型的有效性。

（2）针对多目标煤与煤矸石形态差异不大、目标图像光线分布不均匀等导致识别不理想的问题，提出了一种基于改进FCN的煤与煤矸石多目标分割模型。首先，选择FCN为基础模型，采用ResNet50为FCN的主干网络，采用膨胀卷积并结合HDC设计准则增大模型感受野，提取目标图像的全局特征。其次，在ResNet50网络中引入通道注意力机制和像素注意力机制，并以串行的形式引入残差结构网络中，使得分割网络分割出的像素级别的区域能够更准确的分配到预定义的语义类别中，提高识别效率。最后，将改进后的模型应用于多目标煤与煤矸石数据集中。实验结果表明，改进后的FCN模型性能指标均有所提升，Pixel Accuracy达97.3%，mean Accuracy达95.2%，mean IOU达90.4%。经过验证，改进后模型的分割效果、识别效果均得到有效提升。

（3）设计并实现了煤矸识别系统。基于改进EfficientNetV2和FCN煤矸识别模型，结合Pytorch和Flask框架设计煤矸识别系统，对输入到平台的煤与煤矸石图像进行识别，同时系统还包含识别记录模块以及论坛分享功能，为煤与煤矸石图像的分析研究提供支持。

       本文提出了两种基于深度学习的煤矸识别模型，解决单目标与多目标场景下出现的煤矸识别问题。同时设计并实现了煤矸识别系统，实现对煤与煤矸石的高效识别。

Coal and gangue sorting is an important step in the coal mining process. The chemical components in the coal gangue mixed with the coal reduce its combustion efficiency and utilisation rate, and produce a large amount of harmful gases that harm the human body and pollute the environment during the combustion process. The conventional approach to coal gangue identification is costly, laborious, and prone to other limitations, rendering it inadequate for the actual needs of coal gangue sorting. With the advent of deep learning technology in diverse fields, the application of neural networks has become a promising avenue for more accurately extracting the intricate details present in images, thereby significantly enhancing the recognition efficiency. Therefore, coal and gangue recognition based on deep learning has become an important means to solve coal and gangue sorting. Nevertheless, there are also issues pertaining to the accuracy of the model recognition and the efficacy of multi-target recognition. In this paper, based on the image classification model and semantic segmentation model, single target classification and multi-target segmentation methods in coal and gangue recognition are studied. The main research work includes:

(1) Aiming to address the multi-scale morphology of coal and gangue in actual production environments, the existing coal selection methods are inadequate for coal and gangue recognition and other issues. To address these limitations, a single-target recognition model of coal and gangue based on an improved EfficientNetV2 is proposed, which combines image data preprocessing and image classification models. Firstly, EfficientNetV2 is selected as the base model, and the data set is expanded using pipeline-based data enhancement methods and image processing techniques before model training. This is done in order to simulate the production environment of coal and gangue under actual working conditions. Secondly, the MBConv and Fused-MBConv modules in the original model are enhanced, and the CAM attention mechanism module and Hardswish activation function are introduced to address the limitations of the original attention mechanism module and activation function. This simultaneously improves the network's recognition accuracy. Finally, by optimizing the hyperparameters of the network structure, the training speed of the network is accelerated while maintaining the accuracy of the model. The improved model is then applied to the coal and gangue dataset after data preprocessing. The experimental results demonstrate that the enhanced EfficientNetV2 coal and gangue recognition model is straightforward to train, rapid in training speed, rapid in convergence, and highly accurate in recognition. Compared to the original model, the recognition accuracy has been enhanced by 3.98% to 98.24%. Furthermore, the training speed has been enhanced, with the inference time of the improved model reduced by 6.6 milliseconds. This is in contrast to the original inference of 39 images per second, which has increased to 52 images per second. This verifies the effectiveness of the improved model.

(2) Aiming to address the issue of inadequate recognition due to subtle morphological differences in multi-target coal and gangue as well as the uneven distribution of light in the target image, a multi-target segmentation model of coal and gangue based on an enhanced FCN is proposed. Firstly, FCN is selected as the base model. ResNet50 is adopted as the backbone network of FCN, and the expansion convolution is used in conjunction with the HDC design criterion to increase the model's sensory field, thereby enabling the extraction of global features from the target image. Secondly, the channel attention mechanism and pixel attention mechanism are introduced into the ResNet50 network and incorporated into the residual structure network in a serial manner. This enables the pixel-level regions segmented by the segmentation network to be assigned to predefined semantic categories with greater accuracy, thereby enhancing the recognition efficiency. Finally, the improved model is applied to the multi-target coal and gangue dataset. The experimental results demonstrate that the performance metrics of the enhanced FCN model have been enhanced, with pixel accuracy reaching 97.3%, mean accuracy reaching 95.2%, and mean IOU reaching 90.4%. This verifies that the segmentation efficacy and recognition efficacy of the enhanced model have been effectively enhanced.

(3) The coal and gangue recognition system was designed and implemented. It is based on the improved EfficientNetV2 and FCN coal and gangue recognition model, which was combined with the PyTorch and Flask frameworks to design the coal and gangue recognition system. This system is capable of recognizing coal and gangue images that are input to the platform. Additionally, it contains a recognition record module and a forum sharing function, which provide support for the analysis and research of coal and gangue images.

       This paper presents two deep learning-based coal and gangue recognition models, which are designed to address the challenge of coal and gangue recognition in both single-target and multi-target scenarios. Additionally, a coal and gangue recognition system is developed and implemented to achieve the efficient recognition of coal and gangue.