Space object detection algorithm based on adaptive feature texture enhancement and receptive field adjustment

With the continuous development of artificial intelligence, the use of deep learning to achieve intelligent space object detection has become a new research trend. Space-based observation platforms are affected by the space environment with many problems such as small scale of space object, large amount of noise, low recognition and little extractable information. To address the above issues, an improved fully convolutional one-stage object detection (FCOS) model based on adaptive feature texture enhancement and receptive field adjustment is proposed. To address the problem of pixel smoothing and detail loss caused by up sampling in convolutional neural networks (CNN), this paper proposes a texture detail enhancement module (TDEM), which is based on sub-pixel convolution to achieve effective scaling of the feature map by automatically learning the interpolation function and enhance the correlation between the pixels of the image while suppressing irrelevant features. In addition, in order to obtain more dense features and appropriate receptive fields, an adaptive receptive field adjustment module (ARFAM) is proposed by using densely connected dilated convolution and attention mechanism to enrich the contextual information around the object and improve the detection capability of the model. This paper constructs the SDM dataset, which contains 6842 images and three categories of satellites, debris, and meteorites. The experimental results on the SDM dataset show that our method achieves the mAP of 73.9%, which illustrates detection performance is significantly better than the mainstream algorithms.