在复杂煤矿环境和光照不均等恶劣条件下，现有的目标检测技术难以满足智能巡检需求。为解决这一问题，提出了一种基于改进卷积神经网络和自注意力机制的目标检测方法。首先构建了基于金字塔结构与注意力机制的特征提取网络；在此基础上，设计了双向特征金字塔网络模块，进一步强化特征提取功能；最后，利用YOLO Head模块进行预测处理。试验结果表明：经过2 398次迭代后，模型的损失最终稳定在0.01左右，达到了理想的loss效果；模型在迭代至1 598次时准确度达到0.95，并在1 845次左右进入稳态，是达到稳态最快的模型之一，整体检测时长为3.2 ms；模型能够提升复杂环境下目标检测的精度和效率。

In complex coal mine environment and poor lighting conditions, the existing target detection technology is difficult to meet the needs of intelligent inspection. To solve this problem, an object detection method based on improved convolutional neural network and self-attention mechanism is proposed. Firstly, a feature extraction network based on pyramid structure and attention mechanism is constructed. On this basis, the bidirectional feature pyramid network module is designed to further strengthen the feature extraction function. Finally, the YOLO Head module is used for prediction processing. The test results show that after 2 398 iterations, the loss of the model is stabilized at about 0.01, and the ideal loss effect is achieved. The accuracy of the model reached 0.95 at 1 598 iterations and entered steady state at about 1 845 iterations, which was one of the fastest models to reach steady state. The overall detection time was 3.2 ms. The model can improve the accuracy and efficiency of target detection in complex environment.