Research on Target Localization Technology based on Depth of field fusion of generative adversarial networks

Target positioning is a crucial step in beam target coupling, but the multi channel imaging system of the beam target coupling sensor struggles to obtain reliable target position information due to its small depth of field and poor illumination. This paper proposes a depth of field fusion algorithm based on generative adversarial networks that can merge multiple partially clear target images into a single image with higher clarity, thereby extending the depth of field and improving the accuracy of target positioning. Specifically, we first use OpticalFlow Net to estimate the optical flow to warp the source image and defocused image to a unified camera view, then generate an image with clear target edges using Fusion Net; finally, we extract the target position parameters through post processing methods such as edge detection to assist in high precision positioning. Additionally, to train our network, we created a multi focus pixel misaligned dataset based on the principle of point spread. Experimental results show that our algorithm outperforms existing advanced methods in both quality and effectiveness, being able to control the target positioning accuracy within 2.7μm, meeting the accuracy requirements.