Optimization of Small Object detection based on Generative Adversarial Networks

Small object detection is one of the fundamental problems in computer vision applications. Existing small object detection techniques usually focus on detecting small objects with multiple scale of features with low efficiency due to high computational cost. In this paper, we investigate small object detection problem based on generative adversarial architecture that utilizes features of small objects. We propose an Optimized Perceptual Generative Adversarial Network (OPGAN) to present more features of small objects. Specifically, the generator of OPGAN learns to present the low-resolution features of the small objects to highly resolved features similar to large objects as input image of the discriminator model. After then, the discriminator of OPGAN computes the generated feature and generates a new perceptual requirement parameter into the model to train the model iteratively. Extensive experiments on the challenging benchmark data sets demonstrate the effectiveness of OPGAN in detecting small objects