With the recent advancement of the Internet of Things (IoT) and edge computing, there’s a surge in demand for effective intelligent data analytics, especially vision analysis. Neural network (NN) partitioning, an edge-cloud collaborative computing technique that divides NN models into distributable segments, offers a promising approach to deploy efficient AI-based intelligent system systems. However, to complete tasks such as after-the-fact video query and manual labeling in the cloud, the servers need to receive both intermediate features and original images, leading to significant communication redundancy. In this paper, we propose an approach to transmit a single data flow to support both image restoration and NN inference tasks. We design a bottleneck unit for fusing and compressing intermediate features and an image restoration module with self-attention blocks. Our proposed framework can be easily adapted to existing NN partitioning systems without modifying the structure of the model. With comprehensive evaluations, our system can reduce the amount of network transmission data by 57.3% within 1% accuracy loss, while introducing negligible overhead on local devices.