基于边缘引导的自校正皮肤检测

Edge Guided Self-correction Skin Detection

皮肤检测作为计算机视觉领域中的研究热点多年来被广泛研究,且仍然是一项具有挑战性的任务。尽管目前的方法在许多常规场景下取得了成功,但仍然存在预测不完整和泛化能力差等问题。针对该问题,提出了一种基于边缘引导的神经网络,并且由大量经过自校正的皮肤检测数据驱动网络训练,实现鲁棒的皮肤检测。首先,提出一种基于多任务学习的网络,对皮肤检测和边缘检测两个任务进行联合优化。进一步,提出边缘注意力模块,将预测所得的边缘检测结果通过该模块重新融合到皮肤检测支路中。最后,提出一种自校正算法,通过借助人体解析任务中的大量低质量数据以增强皮肤检测模型的泛化能力。通过自校正算法对带噪声标签的优化,逐步消除使用带噪声标签进行监督训练的副作用。实验结果表明,所提皮肤检测方法优于现有的其他方法。

Skin detection has been a widely studied computer vision topic for many years,whereas remains a challenging task.Previous methods celebrate their success in various ordinary scenarios but still suffer from fragmentary prediction and poor generalization.To address this issue,this paper proposes an edge guided network driven by a massive self-corrected skin detection dataset for robust skin detection.To be specific,a multi-task learning based network which conducts skin detection and edge detection jointly is proposed.The predicted edge map is further converged to the skin detection stream via an edge attention module.Meanwhile,to engage a large-scale of low-quality data from the human parsing task to strengthen the generalization of the network,a self-correction algorithm is adapted to prune the side effect of supervised by noisy labels with continuously polishing up those defects during the training process.Experimental results indicate that the proposed method outperforms the state-of-the-art in skin detection.