给出一种基于轻量化卷积神经网络的空间非合作目标局部特征检测网络，即NCDN模型。在SSD模型中引入特征融合策略以适应不同距离下的检测需求，提高模型对图像尺度变换引起局部特征分辨率降低的鲁棒性；并采用不同压缩比例对MobileNetV2内部卷积通道数量做压缩，从而得到轻量化特征提取网络；对SPEED数据集进行局部特征标注与训练以验证NCDN适用的距离范围。实验结果表明，该模型能够在45 m内距离范围保证mAP达到0.90，同时通道压缩节省75%计算量后模型精度损失仅为5%。满足在轨检测精度和计算量需求。

The Non-cooperative Detection Network(NCDN) model is a kind of local feature detection network based on lightweight convolution neural network. In SSD model, the feature fusion strategy was introduced to meet the detection requirements at different distances, and the robustness of the model to the reduction of local feature resolution caused by image scale transformation was improved; the number of convolution channels in mobilenetv2 was compressed with different compression ratios to obtain lightweight feature extraction network; local feature labeling and training of speed data were set to verify the applicable distance range of NCDN. The experimental results show that the mAP of the model can reach 0.90 within 45 m, and the accuracy loss of the model is only 5% after saving 75% of the calculation amount in channel compression. It meets the requirements of on orbit detection accuracy and calculation amount.