煤矿作为一个高危行业，其生产安全问题一直以来都备受高度重视。煤矿井下安全事故频发，安全生产的形势很严峻。在井下作业环境的监控摄像头覆盖区域大，但由于拍摄距离较长，画面中井下目标像素较小，导致对井下小目标检测中存在许多问题。一方面，井下光线条件通常较差且目标距离监控摄像头较远，使得其在图像中所占的像素较少，导致井下小目标特征信息提取的不准确；另一方面，井下复杂场景中边缘检测设备计算能力受限，煤矿井下实时拍摄的视频数据量巨大，导致网络模型的参数量和计算量变得非常复杂，在煤矿井下环境中检测速度慢。因此，本文针对井下小目标检测方法及其实时性展开研究，工作内容如下：

（1）构建了井下小目标数据集。针对当前井下小目标数据集欠缺的问题，本文通过收集方式获取井下小目标图像，构建一个多场景的井下小目标数据集。然后采用数据扩充和数据标注等方法进行数据预处理。最后，利用去雾化、锐化等图像处理方法提高本文数据集的质量。数据集的构建为后续研究提供了重要基础，利用该数据集进行井下小目标检测算法的训练和评估。

（2）针对煤矿井下环境中小目标特征信息提取不准确的问题，提出了基于YOLOv7-SE的井下小目标检测方法。首先，融合模拟退火算法和k-means++聚类算法，优化井下小目标初始锚框值的估计。其次，本文在网络模型中增加新的井下小目标检测层，减少细粒度井下小目标特征丢失。同时在ELAN模块之后嵌入双层注意力机制，强化井下小目标特征表示。最后，引入新的井下小目标度量指标，以解决IoU对于井下小目标偏差非常敏感的问题。实验结果表明，YOLOv7-SE 网络模型井下小目标平均精度mAP值为68.47%，与YOLOv7网络模型相比，mAP值提高了8.12%，可以有效增强井下小目标特征提取。

（3）针对井下复杂场景中因边缘检测设备计算能力受限，影响井下小目标检测速度的问题，提出了基于YOLOv7-tiny-PB的井下轻量化小目标检测方法。该方法基于FasterNet模型进行井下小目标的特征提取，通过改进ELAN模块，减少卷积计算规模，加快模型对井下小目标的推理速度。其次，设计一种参数量更少的BiFPN-s网络，高效捕捉井下小目标特征信息。最后，结合通道剪枝和知识蒸馏方法进一步压缩模型，以便获得轻量化模型。实验结果表明，本文所提方法较基模型以及其他轻量模型参数量大幅度下降，参数量仅为基模型的25.42%，计算量为基模型的49.49%，模型大小仅为2.7M，可以在减小模型参数规模的同时，快速检测井下小目标。

（4）面对井下小目标检测，本文设计并实现了一套基于煤矿井下的小目标检测系统。综合上述提出的轻量化小目标检测方法，对检测结果在本系统进行可视化展示。该系统是基于PyQt5框架实现目标检测模型的界面化，能够根据用户预设的参数得到检测图像中的目标位置和目标类别。

As a high-risk industry, the safety of coal mine has always been highly valued. There are frequent safety accidents in coal mine, and the situation of safety is very serious. The monitoring camera in the underground working environment covers a large area, but due to the long shooting distance, the underground target pixels in the picture are small, resulting in many problems in the detection of small underground targets. On the one hand, the underground light conditions are usually poor and the target is far away from the surveillance camera, which makes it occupy less pixels in the image, resulting in inaccurate extraction of small target feature information in the underground. On the other hand, the calculation ability of edge detection equipment in complex underground scenes is limited, and the amount of video data captured in real time in coal mine is huge, which leads to the complexity of the parameters and calculation of the network model, and the detection speed is slow in the underground environment of coal mine. Therefore, this thesis studies the underground small target detection method and its real-time performance. The work content is as follows :

(1)A downhole small target data set is build. In view of the lack of current downhole small target data sets, this paper obtains downhole small target images by collecting methods, and constructs a multi-scenario downhole small target data set. Then, data preprocessing is carried out by data expansion and data annotation. Finally, image processing methods such as de-atomization and sharpening are used to improve the quality of the data set in this thesis. The construction of the data set provides an important basis for subsequent research. The data set is used to train and evaluate the underground small target detection algorithm.

(2)Aiming at the problem of inaccurate extraction of small target feature information in coal mine underground environment, a small target detection method based on YOLOv7-SE is proposed. Firstly, the simulated annealing algorithm and k-means++ clustering algorithm are combined to optimize the estimation of the initial anchor frame value of the underground small target. Secondly, this paper adds a new underground small target detection layer to the network model to reduce the loss of fine-grained underground small target features. At the same time, a double-layer attention mechanism is embedded after the ELAN module to strengthen the feature representation of underground small targets. Finally, a new measurement index of underground small target is introduced to solve the problem that IoU is very sensitive to the deviation of underground small target. The experimental results show that the average accuracy mAP value of the YOLOv7-SE network model is 68.47%. Compared with the YOLOv7 network model, the mAP value is increased by 8.12%, which can effectively enhance the feature extraction of underground small targets.

(3)Aiming at the problem that the calculation ability of edge detection equipment is limited in complex underground scenes, which affects the detection speed of small underground targets, a lightweight small underground target detection method based on YOLOv7-tiny-PB is proposed. This method is based on the FasterNet model to extract the features of small underground targets. By improving the ELAN module, the scale of convolution calculation is reduced, and the inference speed of the model for small underground targets is accelerated. Secondly, a BiFPN-s network with fewer parameters is designed to efficiently capture the feature information of small underground targets. Finally, the model is further compressed by combining channel pruning and knowledge distillation methods to obtain a lightweight model. The experimental results show that the proposed method has a significant decrease in the number of parameters compared with the base model and other lightweight models. The number of parameters is only 25.42 % of the base model, the calculation amount is 49.49 % of the base model, and the model size is only 2.7 M. It can quickly detect small underground targets while reducing the scale of model parameters.

(4)In the face of underground small target detection, this thesis designs a set of small target detection system based on coal mine underground. Based on the above proposed lightweight small target detection method, the detection results are visually displayed in the system. The system is based on the PyQt5 framework to realize the interface of the target detection model, and can obtain the target position and target category in the detected image according to the user 's preset parameters.