Optimizing Adaptive Image Anomaly Detection Strategies Based on Reinforcement Learning
Ying Cao, Zhengyang Li, Yuqing Wang, Zhen Tian · 2025
Image anomaly detection is a critical technology in the fields of computer vision and quality control. To address the problem of existing methods being prone to local optima, this study proposes a novel adaptive optimization strategy based on deep reinforcement learning. This strategy takes convolutional neural network features as input states, anomaly classification as actions, and approximates the policy function through a deep Q-network. It adjusts the strategy based on feedback during each interaction with the environment to adapt to different anomaly patterns. Experimental results demonstrate that this method outperforms current state-of-the-art methods in various metrics on publicly available datasets, enhancing the ability to detect various types of anomalies. This research provides an effective reinforcement learning optimization strategy for anomaly detection tasks, validating its adaptability and generalization and offering new insights for the development of this field.