Research On One-Stage Object Detection Algorithms Based On Generative Adversarial Network

With the rapid development of deep learning,deep Convolutional Neural Networks have achieved rapid development in the field of computer vision.As a mainstay in the field of computer vision,object detection technology is a challenging task,and it is widely used in unmanned driving,surveillance security and other fields.The currently proposed object detection algorithm has poor detection accuracy for deformed occlusion objects and small objects,and is prone to misdetection and missing detection.Achieving a balance between detection accuracy and speed of object detection algorithms has become a current research hotspot.This paper analyzes and improves the shortcomings of the one-stage object detection algorithm YOLOv3 in the object detection algorithm.Compared with the YOLOv2 algorithm,the YOLOv3 algorithm improves the detection accuracy of small objects,but the detection accuracy of medium and large objects has decreased,and the detection accuracy of twisted and deformed objects is still insufficient.In order to improve the shortcomings of the YOLOv3 algorithm:(1)To further improve the detection performance of small objects and medium and large objects,the GA-YONET algorithm proposed in this paper uses the super-resolution of generating adversarial networks(Generative Adversarial Networks)Operation,and proposed a gate convolution mechanism and a channel attention mechanism to better locate small objects in feature maps.On the one hand,it can improve the detection accuracy of small objects;on the other hand,it can also be used as a data enhancement operation.(2)To improve the object detection performance under different occlusion degrees,this paper uses the GA-YONET algorithm to generate a confrontation network to add a mask to the feature value of the feature map,and the feature value under the mask is set to 0.The object detection for twisting and deforming uses horizontal or vertical stretching and rotating operations on the feature map.(3)Improvement of the loss function:In order to better improve the detection accuracy of the object,this paper first of all the commonly used loss functions such as KL(Kullback-Leibler divergences)divergence and JS(Jensen-Shannon divergences)divergence Mathematical thoughts were analyzed and it was found that KL divergence and JS divergence had defects when used as a loss function for deep learning.Therefore,this paper uses the distance-based Wasserstein Distance as the loss function.The advantage of the Wasserstein Distance loss function is that this function can better describe the true distance between the two divisions,and in order to reduce the difference in the number of positive and negative samples in the training phase,this article adds a Gaussian penalty term to the Wasserstein Distance loss function.By modifying the weights,the deep neural network focuses on the loss of classification errors.In order to prevent the gradient explosion,the weight update operation is optimized,and the Laplacian smoothing is used to reduce the gradient return value when the gradient is returned to 0.By improving the shortcomings of the above YOLOv3 algorithm,the GA-YONET one-stage object detection algorithm improves the detection accuracy of small objects and occlusion deformation objects,and has different degrees of effect improvement on the MSCOCO 2012 data set.The YONET algorithm was transplanted into the unmanned vehicle and the effect demonstration was also performed in outdoor real scenes,including the detection of night and snow scenes,and achieved good detection results.Confirmed the effectiveness of this paper to improve the YOLOv3 algorithm.