An Enhanced Algorithm for Object Detection Based on Generative Adversarial Structure
The performance of object detection networks is often limited by the depth of the feature extraction network.Increasing network parameters may yield limited improvements in the detection system's performance.Additional careful designs of network details are necessary,but they can significantly increase training difficulty.In this paper,generative adversarial networks are used as a method to further enhance the feature extraction capability of the network.In the normal architecture,by leveraging generative adversarial networks(GANs),it becomes possible to approximate the target distribution of a given task.This approach seeks a"near correct answer"by iteratively optimizing a non-convex game with continuous high-dimensional parameters.The generator and discriminator within the GAN framework strive to achieve a Nash equilibrium,resulting in an effective solution for the task at hand.In GANs,gradient descent is commonly employed to handle losses on both the generator and discriminator sides.In this paper,it is experimentally demonstrated that feature-highlighted images have similar feature distributions as unprocessed images,while the evolution of such feature distributions exhibits a continuous and learnable change as the degree of feature highlighting changes.Therefore,this paper introduces a new object detection method using generative adversarial training,which utilizes the ability of generative adversarial networks to fit feature distributions to enhance our object detection network.Our approach focuses on minimizing the EM distance(Wasserstein distance)of the feature distri-bution,using features acquired with technically processed images as a benchmark to create a target distribution for the generative adversarial network.The features obtained from the original images will be considered as false information in generative adversarial,and the process of adver-sarial training will continuously improve the feature extraction capability of the network to obtain more realistic features,thus improving the target detection capability.Simultaneously,due to enhanced image features,the training of GAN(Generative Adversarial Network)yields a feature distribution that exceeds that of the original dataset,which allows additional gains to be obtained more easily than the usual training methods.A new loss function is also added during adversarial training to ensure steady improvement of the detector by constantly checking the object detection performance of the network.A comparative experiment conducted with the original CenterNet network on MS COCO(Microsoft Common Objects in COntext)2017 reveals that the generative adversarial training method significantly improves the average precision for most of the examined backbone networks,while ensuring that there is no increase in the inference complexity of the network.Among the four backbone networks employed in the experiments,the mean improvement in network AP(Average Precision)values ranged from 0.3 to 0.9,demonstrating their success with minimal training efforts.Moreover,none of the four backbone networks experienced an increase in network parameters during inference.Experimental results indicate that the proposed architecture effectively enhances the network's feature extraction capability without compromising speed during inference.
万方数据
维普
