Sun Jiawei,∗ Hao Jia,∗ Li Yimin,∗ and Yan Yan∗
Open-world incremental object detection, industrial inspection, clustering learning, transformer-based object detection network ∗ School of Mechanical Engineering, Beijing Institute of Technology, Beijing, China; e-mail: {895984926, liyiminjason}@qq.com; [email protected]; [email protected] Corresponding author: Hao Jia
Industrial safety inspection is crucial task to ensure the safety of the production environment and the normal operation of production equipment. With the continuous development of artificial intelligence, industrial visual detection technology has attracted widespread attention as a non-contact automated inspection method. However, the actual inspection environment is complex and ever- changing, and the categories and characteristics of the detection targets change dynamically, which requires the visual detection algorithm to learn and train based on the known category data while being able to recognise known and unknown categories incrementally. Besides, existing static visual detection algorithms cannot meet these requirements, such as insufficient generalisation ability of models, weak adaptability to new data and unknown targets, catastrophic forgetting of learned known categories caused by learning unknown category data, and so on. We observe that when humans identify an object, they subconsciously focus on the foreground object firstly, then they recognise the details of each object, rather than simultaneously locating and identifying an object. Inspired by this idea, this paper focuses on the research on industrial visual detection algorithms based on a detector–recogniser decoupling transformer detection network, which can automatically identify and detect objects of new categories without affecting the detection performance of known category objects. Meanwhile, we propose a self-adaptive pseudo-labelling mechanism which combines the attention-driven with knowledge-driven pseudo-labels mechanism and self-adaptively generates pseudo-labels for unknown objects robustly to continuously iterate and optimise the inspection model. Extensive experiments on our self-collected dataset and MS-COCO demonstrate that our model outperforms the other state-of-the-art methods.
Important Links:
Go Back
