“The research team proposed a deep separable convolutional neural network model for target recognition on micro nano satellite in orbit computing platforms. This model reduces the depth and complexity of the overall network structure by improving the Yolov4 network structure, and uses separable convolution structures to reduce the number of model parameters. Meanwhile, by merging the convolutional layer with the Batch Normalization layer, the forward inference speed is further accelerated. In addition, the research team also drew inspiration from the Focal loss function and improved the loss function of the object detection network to alleviate the problem of imbalanced foreground and background sample ratios. Compared with the original Yolov4 network model, the new model reduces the number of parameters by about 7 times and FLOPs by about 30 times while ensuring a recognition accuracy of 94.09%. In addition, the algorithm performance of the new model was further validated through comparative experiments with cutting-edge network models such as Yolo series, SSD, MobileNet, CenterNet, etc. This research achievement provides theoretical support for achieving in orbit target recognition and filtering useless data, which helps to promote the technological progress and application expansion of micro nano satellite in orbit computing platforms.”