输电线路作为电力传输的纽带，提供了可靠的电力供应，随着西电东送、特高压线路等一大批国家战略电能的部署，使得输电线路规模动辄横跨数千公里。由于其所处环境为野外，导线极易遭受异物附着，为保障电力的安全运行，必须加强巡检排查安全隐患。目前有人工巡检和无人机巡检两种方式，相较于传统人工巡检效率低、实时性低和危险性高等问题，无人机巡检技术展现出无可比拟的优势。无人机最重要的一种感知手段是前端摄像头，结合深度学习可以实现对感知图片自动筛查异物的目的。为了实现快速精准的检测，本文以YOLOv8算法为核心对航拍图像中的4种异物为目标进行检测。主要研究工作如下：

(1) 优化样本质量并完成标注工作，首先构建包含鸟巢、风筝、气球、塑料制品四类常见异物数据集，为后续模型训练提供可靠依据。其次，针对多尺度异物辨识精确度欠佳问题，提出一种自适应特征金字塔模块(AFPN)代替原有特征网络，联结异物内在的初级与高级视觉特征，使各层级特征映射具备自适应权重输出的能力，识别精度相比YOLOv8增加了1.8%；再在主干网络末端添加GAM注意力机制，弥补了跨维度特征信息的损失的不足，识别精度增加了3.2%；接着将CIoU损失函数替换成EIoU损失函数，增强对远景目标的检测，同时利用Soft-NMS算法减少漏检率，提升对输电线路小目标的检测能力，使网络性能提升了4.4%。通过在算法中使用AFPN自适应特征融合模块、融入GAM注意力机制、采用EIoU损失函数与Soft-NMS算法，与Faster RCNN、YOLOv5、YOLOv7、YOLOX、YOLOv8和本文算法对数据集进行异物检测实验对比，验证本文算法平均精度高达97.6%。

(2) 在训练好的模型中对CBS归一层通道进行了剪枝量化操作，改进主干网络对应的特征图通道数量，再在主干网融入轻量级的CA注意力模块，使模型达到平衡速度与精度的要求，模型参数量为18.64MB，相比于原始网络降低了25.11MB，模型精确度为95.4%，推理速度达到了17.53ms。最后将剪枝后的模型导入到输电线路异物检测界面，设计GUI界面使用户可以进行交互，完成检测的执行与监控能够极大的提升电网巡检人员的测试效率，具有很高的工业应用价值。

Transmission lines as a link of power transmission, provides a reliable power supply, with the west to east, extra high voltage lines and a large number of national strategic power deployment, so that the scale of transmission lines can easily span thousands of kilometers. Because of the environment in which it is located for the field, the conductor is very easy to suffer from foreign body invasion, in order to protect the safe operation of electricity, must strengthen the inspection to check the safety hazards. At present, there are two ways of manual inspection and drone inspection, compared with the traditional manual inspection efficiency is low, real-time and dangerous problems, drone inspection technology shows incomparable advantages. One of the most important perception means of UAV is the front-end camera, in which deep learning can be applied to realize the purpose of automatic screening of foreign objects in the perception picture. In order to realize fast and accurate detection, this paper takes YOLOv8 algorithm as the core to detect four kinds of foreign objects as targets in aerial images. The main research work is as follows:

(1) Optimize the quality of the samples and complete the annotation work, and construct a dataset containing four types of common foreign objects, including bird's nests, kites, balloons, and plastic products, to provide a reliable basis for subsequent model training. Second, to address the problem of poor multi-scale foreign object recognition accuracy, an adaptive feature pyramid module (AFPN) is proposed instead of the original feature network, linking the intrinsic primary and advanced visual features of the foreign object, so that the feature mapping at each level has the ability of adaptive weight output, and the recognition accuracy is increased by 1.8% compared with that of YOLOv8; a GAM attention mechanism is added to the 9th and 10th layers of the backbone network which compensates for the lack of loss of cross-dimensional feature information, and the recognition accuracy increases by 3.2%; then the CIoU loss function is replaced with the EIoU loss function to enhance the detection of farsighted targets, and the Soft-NMS algorithm is utilized to reduce the leakage detection rate and improve the detection capability of small targets on transmission lines, which improves the network performance by 4.4%. By using the AFPN adaptive feature fusion module in the algorithm, incorporating the GAM attention mechanism, adopting the EIoU loss function with the Soft-NMS algorithm, and comparing the foreign object detection experiments on the dataset using Faster RCNN, YOLOv5, YOLOv7, YOLOX, YOLOv8 and the algorithms in this paper, we validate that the algorithms in this paper have an average accuracy of up to 97.6%.

(2) In the trained model of the CBS to a layer of channel pruning quantization operations, improve the backbone network corresponding to the number of channels of the feature map, and then in the backbone network into the lightweight CA attention module, so that the model to achieve a balance between speed and accuracy requirements, compared to the original network model size reduced by 2.3 times, the model accuracy of 95.4%, the inference speed of 17.53 ms. Finally, the model will be imported into the transmission line multi-target detection interface, the design of the GUI interface so that the user can interact with the completion of the inspection and monitoring can greatly improve the efficiency of the test of power grid inspectors. The pruned model is imported into the transmission line multi-target detection interface, and the GUI interface is designed so that the user can interact with it, and the execution and monitoring of the detection can greatly improve the testing efficiency of the grid inspectors, which has a high industrial application value.