矿井胶带火灾作为煤矿火灾的主要形式之一，其带来的损失是无法估量的。因此对矿井胶带火灾实现智能化识别对保障矿井工作人员的生命安全具有极其重要的意义。随着矿井自动化水平的提升，传统的胶带火灾识别方法已经无法满足矿井当下的智能化需求，而基于深度学习的火灾视频图像识别方法具有响应快、准确度高等多方面的优势，将其应用到矿井胶带火灾识别，对于提高矿井的安全系数具有一定的实用价值。本文针对深度学习技术应用到矿井胶带火灾识别存在的问题进行如下研究：

针对矿井下的复杂环境对胶带火灾视频图像造成的颜色失真、色彩度低等问题，提出一种基于K-means实现的暗通道去雾改进算法对图像进行处理，提高图像的可读性信息，使得图像的特征信息更加丰富。同时通过帧差法与高斯混合模型融合的算法对井下动态烟火信息进行提取，降低与烟火目标相似的静态非目标所造成的干扰，提高胶带火灾的识别精度。

针对AlexNet模型的胶带火灾识别准确率低、计算参数复杂等问题，采用小尺寸卷积核对AlexNet模型前两层的大尺寸卷积核进行替换，来获取更加丰富的特征信息，提高模型的识别精度和训练速度。同时采用Mish函数替换原始模型中的Relu激活函数，提高模型的泛化能力，根据实验结果可以得到改进后的AlexNet胶带火灾识别模型的识别准确率达到88.3%，较改进前提升了5.9%，模型的性能得到了有效的改善。

针对YOLOv4（You Only Look Once，YOLO）胶带火灾识别模型的准确率低且训练时间长等问题，本文结合动态注意力机制和群组归一化对原模型进行改进，避免特征信息丢失，提高模型的识别精度。同时采用深度可分离卷积网络及随机池化对原算法模型的跨阶段局部网络（Cross Stage Partial，CSP）及空间金字塔池化（Spatial pyramid pooling，SPP）模块进行优化，经实验验证改进后的F-YOLOv4胶带火灾识别模型的平均检测精度达到97.3%左右，较原始模型提升了7.85%，该模型对矿井胶带火灾的识别具有一定的实用价值。

As one of the main forms of coal mine fires, mine belt fires cause immeasurable losses. Therefore, the realization of intelligent identification of mine belt fires is of great significance to ensure the safety of mine workers. With the improvement of the level of mine automation, the traditional belt fire identification method has been unable to meet the current intelligent needs of the mine, and the fire video image recognition method based on deep learning has the advantages of fast response and high accuracy, and its application to mine tape fire recognition has certain practical value for improving the safety factor of mines. This paper studies the problems existing in the application of deep learning technology to mine belt fire identification are studied as follows:

Aiming at the problems of color distortion and low chromaticity caused by the complex environment in the mine, the proposition about dark channel dehazing algorithm is improved based on K-means which is to process the image and increase the readability of the image. Then, the feature of image will be more rich. At the same time, the dynamic pyrotechnic information in the well is extracted by the fusion algorithm of the frame difference method and the Gaussian mixture model, which reduces the interference caused by the static non-target similar to the pyrotechnic target and improves the identification accuracy of the belt fire.

In view of the problems of low belt fire recognition accuracy and complex calculation parameters of the AlexNet model, small-sized convolution kernels are used to replace the large-sized convolution kernels in the first two layers of the AlexNet model to obtain richer feature information and improve the recognition accuracy of the model and training speed. Meanwhile, the Relu activation function

replaced by the Mish function. Then, the generalization ability of the model is improved. According to the experimental results, the recognition accuracy of the improved AlexNet belt fire recognition model can reach 88.3%, which is 5.9% higher than that before the improvement. Performance has been effectively improved.

In view of the low accuracy and long training time of the YOLOv4 (You Only Look Once，YOLO) belt fire recognition model, this paper combines the dynamic attention mechanism and group normalization to improve the original model to avoid losing information and increased the model’s recognition accuracy. At the same time, the Cross stage partial network (CSP) and Spatial pyramid pooling (SPP) modules of the original algorithm model are optimized by using the deep separable convolution network and random pooling.The map of the improved F-YOLOv4 belt fire identification model has reached about 97.3% after experimental verification, which is 7.85% higher than the original model, the model has certain practical value for the identification of mine belt fire.