Deep Semantic Coding for Wireless Image Retrieval

We address the image retrieval problem for a wireless system including an edge server and an edge device. The query image is first compressed by the edge device, and then transmitted into wireless channel, while the edge server retrieves the received image. Different from conventional schemes directly compressing features via unsupervised learning regardless of the database semantic distribution, we design a deep semantic coding (DSC) scheme by integrating the inverted semantic index structure of the database into the coding process, which can utilize the prior semantic information of the database to reduce the bandwidth. We extract the feature vectors from the images via a convolutional neural network and generate the semantic guided code head, which is followed by the product quantization. The experimental results verify the effectiveness of the DSC scheme in reducing the bandwidth as well as improving the performance of wireless image retrieval.