Image Coding for Object Recognition Tasks Based on Contour Feature Learning With Flexible Object Selection

The consumption of image data by machines is rapidly increasing due to the growing adoption of image recognition technologies. This trend has accelerated research in image compression techniques tailored for machine processing. This emerging field, known as Image Coding for Machines (ICM), has gained significant attention in recent years. In particular, ICM is increasingly seen as essential for collaborative systems between edge devices and cloud AI. Since large AI models are challenging to deploy on edge devices, cloud AI services are made available to edge users, who can utilize them by transmitting images to the cloud. Cloud AI is expected to handle various tasks, including image generation and image recognition, with the latter being especially valuable for video and image analysis. Given its utility, image recognition models are anticipated to replace human analysts in applications such as farm and traffic monitoring. Moreover, since recognition models require only a small fraction of the total image data, developing specialized image compression methods for recognition can significantly enhance communication efficiency. However, applying conventional ICM methods to edge-cloud systems presents challenges, such as increased computational load on edge devices and limited versatility. In this paper, we address these challenges by proposing two novel image compression methods—SA-ICM and ST-ICM—designed for recognition models. These methods focus on preserving object contours within images while maintaining compatibility with various recognition models, without adding computational overhead to edge devices. Through experimental evaluations, we demonstrate the versatility and effectiveness of our proposed methods by comparing them with conventional approaches.