Buckle Pose Estimation Using a Generative Adversarial Network

The buckle before the lens coating is still typically disassembled manually. difference between and background small, while that buckles large. This mechanical disassembly can also damage lens. Therefore, it important to estimate pose with high accuracy. paper proposes a estimation method based on generative adversarial network. An edge extraction model designed segmentation network as generator. Spatial attention added discriminator help better distinguish generated real graphs. generator thus generates delicate external contours center lines from discriminator. rectangle least square methods are used determine position deflection angle of buckle, respectively. point accuracies test datasets 99.5% 99.3%, pixel error distance absolute horizontal line within 7.36 pixels 1.98°, achieves highest compared Hed, RCF, DexiNed, PidiNet. It meet practical requirements boost production efficiency coatings.