IMAGE GENERATION METHOD FOR OBJECT DETECTION IN CONTOUR CLEAR SCENE

Abstract

Correct detection of blurred objects has always been a difficult problem for image detection and recognition. Most of the existing methods are to identify fuzzy objects through extensive training by constructing new algorithms. In this paper, we provide a new method to generate semantic layouts by dividing the image of a blurred object according to the contour of the object, then set different labels on the occlusion object, and then generate photographic images using the CRN network. Use the classification network to effectively detect the most likely objects. In more details, our method of synthesizing image is not like traditional methods relying on adversarial training to synthesizing photographic images. Our method uses a single feedforward network, trained end-to-end with a direct regression objective, as a rendering engine, use the 2D semantic specification of the scene to generate the corresponding photographic image. The experimental results on the Cityscapes data set shows that our method can generate the high quality images that is as good as the state of art method while it solves the problem of the awareness of the objects.