To address the challenge of feature-level fusion in sequence image-based space target recognition，this paper proposes a method combining convolutional neural networks (CNNs) and recurrent neural networks (RNNs) with model improvements.In view of the problem of how to use a single image as a sequence feature input，the CNN is modified，in which the feature maps are used as the sequential inputs.In view of the problem of how the sequence features map to the target categories，the long short-term memory (LSTM) network is modified，in which the output layer is enhanced with a new fully connected layer to predict the target categories.Training with the Gaussian noise levels of 0.001~0.006 and testing at 0.007~0.010 achieve a mean average precision (mAP) improvement from 90.7% to 99.16%.Under different postural conditions，the mAP reaches 94.71%.The model has only 283.0 M parameters，effectively addressing the limitations of result-level fusion in existing methods.