Abstract While bringing convenience to our lives, the widespread application of image and video processing technology has recently presented particular privacy issues. Most current mainstream intelligent algorithms rely on the detailed content of images and videos, which significantly increases the risk of personal and private data being leaked. To mitigate this problem, in this paper we use a multi-layer compressed sensing (multi-CS) coding system to degrade the quality of images so their content is not visually discernible. In contrast to other privacy-protection strategies, multi-CS coding protects the content of images and videos while ensuring that intelligent recognition is still possible. A visual privacy protection (VPP) evaluation algorithm is proposed, and this is found to correspond better with subjective evaluation than general image quality-evaluation algorithms. We also propose a statistical modeling method to balance the relationship between the VPP level and the accuracy of intelligent recognition, taking human-pose recognition as an example. Experimental results show the effectiveness and feasibility of the proposed modeling approach.