Object Detection Method In Complicated Environment Based_on Generative Adversarial Networks

Object detection is one of the important research directions in the field of computer vision.With the continuous development of deep learning and computer vision,object detection technology has played a great role in many fields,for example,intelligent monitoring,human-computer interaction,medical diagnosis and military observation.However,in the practical application of object detection,there is many kinds of target to be detected and they are highly susceptible to some factors such as light background,occlusion and deformation.Therefore,it is very challenging to detect objects accurately in real complicated environments.In this paper,the object detection technology based on the generative adversarial networks was studied,including anti-occlusion object detection method,anti-transformation object detection method as well as the joint detection method between anti-occlusion and anti-transformation.The main content and contributions are as follows.1.Aiming at the problem that object images are easily occluded in the real environment,an anti-occlusion detection algorithm is proposed to catch targets from being obscured through using the generative adversarial network.From the data-driven point of view,the adversary with generative occlusion object was constructed for the object detection model,which guide the confrontation between the occlusion generator and the detection model,to expand the occlusion object dataset.Thus,the anti-occlusion performance of the object detection model is improved.2.In view of the low recognition rate of abnormal transformation in the real environment,an anti-transformer network based on the generative adversarial network is proposed.The model can transform objects of images by using spatial transformation algorithm to simulate the transformation interference,so as to realize the confrontation process between the generator and the detection model,promoting the learning efficiently of the object detection model and improving the object detection accuracy of the model.3.We attempted to combine the spatial dropout network with the spatial transformer network into a single network,from the image feature level,achieving a more efficient and more general type of model for object detection under the complex environment.Since the high-level image feature space possess stronger representation ability than the pixel space,this method can further optimize the network structure,reduce the overall cost,and improve the performance of the object detection model at the same time.