Research On Task-driven Deep Image Compressive Sensing

Deep image compressive sensing mainly includes two research contents:sampling matrix design and reconstruction network construction.The quality and speed of image reconstruction are improved by adaptive measurement matrix and optimized reconstruction network.However,when the generated reconstructed images are faced with advanced visual tasks,the distorted reconstructed images interference with the prediction accuracy of the network.Due to the large gap between low-level reconstruction task and high-level prediction task,improving reconstruction quality or prediction accuracy respectively cannot balance the difference.In this paper,the gap between low and advanced tasks is narrowed through end-to-end joint optimization of image reconstruction and task prediction processes.The task prediction ability of reconstruction results can be effectively improved under the condition that the reconstruction quality is guaranteed,and the restored images have rich detail texture.（1）To solve the problem of impaired recognition accuracy of reconstructed images,a compressive sensing network AT-G-₆)optimized by gradient information of target classifier is proposed.The classification loss of pre-trained classifier is added into the image compression and reconstruction network,and the reconstruction and recognition process is optimized end-to-end,so that the low and advanced visual tasks could be combined and adapted to each other.So the compressed and reconstructed image with extensive recognition characteristics is generated.Under the condition that the reconstruction quality is guaranteed,the recognition accuracy is better than that of the compressed sensing network optimized only in pixel space.In addition,the compressed and reconstructed image generated with the gradient information of the known classifier also has an improvement effect on the unknown classifier,with good portability and versatility as well.（2）Further explore the compressed sensing image generation process for unknown classifiers,and design a GAN based compressed sensing network AT-G.When the structure and parameters of the target network are unknown,the discriminant network and category prediction loss are used to make semantic judgment on the generated results,so that the reconstructed images have more perceptual richness.The experimental results show that the reconstructed images generated by AT-G network can be classified by the same categorizer with higher recognition accuracy than the compressed sensing network which only aims at image quality optimization.Moreover,the reconstruction results contain rich texture details and have better visual satisfaction.（3）The compressed sensing network based on gradient information is extended to more visual tasks,and the image compression and reconstruction process oriented to semantic segmentation,human body part segmentation,saliency estimation and other tasks is further studied.A multi task image reconstruction network TD-G-₆)is designed to improve the prediction performance of task-driven reconstructed images.The compressed sensing network driven by advanced visual task can further promote the improvement of reconstruction quality to a certain extent,the edge contour of segmentation result is clearer,and less affected by background,lighting conditions and other factors,and also possesses good adaptability.