Differential-Evolution-Based Generative Adversarial Networks for Edge Detection

Since objects in natural scenarios possess various scales and aspect ratios, learning the rich edge information is very critical for vision-based tasks. Conventional generative adversarial networks (GANs) based methods for edge detection don't perform so well due to model collapse. In order to capture rich edge information and avoid model collapse as much as possible, we consider the learning of GANs as an evolutionary optimization and propose a novel method termed as differential-evolution-based generative adversarial networks (DEGAN) for richer edge detection. In particular, built upon GANs structure, we introduce an improved differential evolutionary algorithm to refine the input of generator, with fitness function evaluated by the discriminator. Experimental results on the well-known BSDS500 and NYUD benchmarks indicate our proposed DEGAN can achieve state-of-the-art performance while retaining a fast speed and validate its simplicity, effectiveness, and efficiency. The high quality of our results on edge detection with proposed DEGAN may promise to make other vision-based tasks work better.