An Adaptive Threshold for the Canny Edge with Weak Label

Edge detection is critical in various computer vision applications such as object recognition, segmentation, and scene understanding. The Canny edge detector remains widely used in traditional methods due to its balance between accuracy and computational efficiency. However, optimal performance depends heavily on carefully selecting three threshold parameters. Prior studies have proposed reinforcement learning-based approaches to automatically determine these parameters. Still, they rely on a supervised edge evaluation network that requires manually annotated high-quality edge labels—a costly and time-consuming prerequisite. This paper proposes a novel approach that eliminates the need for manually labeled data by introducing a weakly supervised reward scheme. We automatically generate weak edge labels from gradient information and use them as pseudo ground truth to compute reward values during the reinforcement learning process. The proposed method leverages an Actor-Critic algorithm to learn adaptive thresholds for the Canny edge detector without explicit supervision. Experimental results demonstrate that our method achieves comparable or superior edge detection performance compared to previous supervised methods and generalizes effectively to unseen datasets.