Lightweight Feature Enhancement Networks Fusing Null Convolution and Attention Mechanisms

Aiming at the comprehensive requirements of model lightweight, high accuracy and small target detection for automatic cargo identification and detection tasks in space station scenarios, this paper proposes an efficient target detection framework, LiteFusionNet, which employs a high-performance lightweight backbone network, EfficientFormerV2, to realize deep semantic feature extraction with low computational cost, taking into account both speed and expressiveness. Expression capability. The Parallel Cavity Convolution and Attention Path Aggregation Network (PACAPN) module is designed, which combines parallel dilation convolution and attention-guided up and down sampling path aggregation mechanism to significantly improve the fusion effect of multi-scale information while maintaining feature consistency. An improved and efficient feature fusion module (SDI) module is introduced, which realizes the dynamic weighted fusion of different scales through the position-adaptive feature interaction mechanism, and effectively enhances the detection capability of small targets and boundary regions. Validated by taking data in the simulated space cargo bay environment, the experimental results show that the map 0.5 of the model reaches 95.4 %, and the map0.5-0.95 improves by 4.6 %, and the parameter count of the model reduces by 41 %, and the FLOPs reduces by 45.7 %. It shows that the model achieves better detection accuracy than the existing original model while maintaining a lightweight structure, demonstrating good potential for engineering deployment.