Improved AOD-Net Dehazing Algorithm for Target Image

The target image plays a crucial role in the visual measurement of the object attitude, and its quality directly affects the accuracy of the object positioning. To improve the quality of target images in adverse weather conditions, such as haze, this paper proposes a serial structure combining U-Net and AOD-Net, along with a newly designed adaptive weighted loss function to enhance the traditional dehazing algorithm of AOD-Net. First, the U-Net, as the network's frontend module, performs multi-scale feature extraction and detail enhancement, preserving edge and texture information. Next, AOD-Net applies fast dehazing to the U-Net-processed image, further improving the global clarity of the image. By leveraging the strengths of both networks through this serial structure, the dehazing accuracy of the target image is significantly improved. Finally, based on the importance of the target region and the level of haze, the loss weight is dynamically adjusted, establishing an adaptive weighted loss function that allows the model to focus more on the restoration of key regions. Experimental results show that the improved AOD-Net algorithm exhibits excellent dehazing performance under various haze conditions. Notably, in complex target images, it effectively enhances image quality and recognition accuracy, with the peak signal-to-noise ratio (PSNR) of the dehazed images improving to 24.61 dB and the structural similarity index (SSIM) increasing to 0.95.