End-to-End Learning Framework Incorporating Image Reconstruction and Recognition Models

Edge computing-based image recognition applications face significant challenges, including increased latency and network load when transmitting large volumes of images to edge servers. To address these issues, this study proposes a novel solution that involves sending compressed data from a front-end device over a network and subsequently reconstructing the images on the server side for recognition purposes. The proposed framework places the image recognition model directly after the image reconstruction model. The reconstruction model is based on a recurrent neural network, and ResNet-18 is used for recognition. This reduction in image quality results in lower recognition performance when using reconstructed images compared with original images. To mitigate this issue, we propose an end-to-end learning framework that jointly optimizes image reconstruction and recognition, specifically optimizing the reconstruction model for the recognition task. The proposed method achieves approximately 99% data compression without degrading classification performance. It improves image quality by 2.1 dB and classification accuracy by 11.2% over the baseline. The proposed method not only significantly improves the quality of reconstructed images without any loss in the image compression rate but also enhances the classification accuracy of these images.