在电力作业场景下，作业人员需要根据不同的作业场景佩戴相应类别的手套。由于作业环境的不确定性引起的手套特征变化，现有算法难以准确地对作业人员所佩戴的手套类别进行有效识别。针对复杂环境下目标鲁棒性特征提取困难的问题，提出一种基于自适应特征提取网络（adaptive feature extraction network, AFEN）的手套细粒度检测算法。首先，AFEN算法在特征提取部分构建自适应特征提取（adaptive feature extraction, AFE）结构，通过自适应调整特征提取模块的大小，以更好地捕捉变化目标的形状特征。其次，算法在特征提取部分构建全局高效注意力机制，帮助模型捕捉不同尺度的特征并更加关注关键特征，提升模型识别的准确性。同时，在检测部分构建小目标检测结构，增强模型对小目标的感知能力，提高模型对小目标的检测效果。最后，AFEN算法使用自适应损失函数，有助于模型更准确地定位目标，进一步提高模型定位精度。较于最优的算法，AFEN算法对手套类别识别准确率达到75.7%，召回率达到73.2%，提升了2.4%，平均精度达到76.3%，提升了2.5%，证明了AFEN算法在手套细粒度检测中的有效性和准确性，并在https：//github.com/wtu1020/ElectricGlove-Dataset公开了手套数据集。

In power operation scenarios, the operators are required to wear gloves of different types depending on the specific task scenario. Due to variations in glove characteristics caused by the uncertainty of the working environment, existing algorithms struggle to effectively identify the type of gloves being worn by workers. To address the challenge of robust feature extraction in complex environments, a fine-grained glove detection algorithm was proposed based on adaptive feature extraction network(AFEN). First, the AFEN algorithm incorporated an adaptive feature extraction (AFE) structure in the feature extraction module, allowing for dynamic adjustments to the size of the feature extraction module to better capture the shape features of varying targets. Second, a global efficient attention mechanism was integrated into the feature extraction module to help the model capture features at different scales and focus more on key features, thereby improving recognition accuracy. Third, a small-object detection structure was incorporated into the detection module to enhance the model's sensitivity to small objects and improved detection performance for such targets. Finally, the AFEN algorithm utilized an adaptive loss function, which enabled the model to more accurately locate targets, further enhancing localization precision. Compared to the state-of-the-art algorithms, the AFEN algorithm achieves a glove type recognition accuracy of 75.7%, a recall rate of 73.2%, an improvement of 2.4%, and an average precision of 76.3%, an improvement of 2.5%, demonstrating its effectiveness and accuracy in fine-grained glove detection, and the glove dataset is released at https://github. com/wtu1020/ElectricGlove-Dataset.