In the field of aerospace research, such as in small celestial body detection, missile guidance, and battlefield reconnaissance, because the target signal is weak, the number of pixels occupied is small, and the target lacks shape structure and texture information, traditional algorithms with manual feature extraction are prone to false alarms, whereas deep learning methods with powerful feature extraction capabilities cannot train tiny targets that lack contour information. In this context, a sliding window sampling training method is adopted, which originates from the idea of nested structures in traditional algorithms based on human visual characteristics. A fully convolutional network using recursive convolutional layers is designed to extend the depth of the network without increasing the training parameters. The multi-branch structure of the network’s parallel convolution structure simulates the multi-scale operation of the traditional algorithm, which can enhance the contrast between the target and the background. Additionally, various loss functions are designed to combat the serious imbalance between positive and negative samples. The results show that the algorithm achieves a better detection performance than the traditional algorithms.