The demand for image dehazing is ever-increasing in image sensor based outdoor systems such as self-driving vehicles, automatic driver assistance, and in highway monitoring and analytics. These applications require dedicated hardware solution to meet high frame-rate and low power constraints. Previously, a few prior based hardware dehazing methods have been presented. However, they produce artifacts in the restored images due to the failure of the underlying assumptions. To address this problem and to meet the stringent requirements, a data-driven image dehazing approach based on convolutional neural network (CNN) and dark channel prior (DCP) is proposed that automatically learns the important features and produces better results. The proposed method is hardware friendly and its hardware implementation is also presented. The design employs few line buffers to store activations and eliminates the requirement of off-chip memory like dynamic random-access memory (DRAM). Field programmable gate array (FPGA) and application specific integrated circuit (AISC) implementations show that the architecture is highly suitable for application scenarios with constrained computational resources, low memory, and tight power budget. The quantitative analysis shows more than 11% average PSNR improvement in image quality on standard datasets as compared to the state-of-the-art hardware methods while consuming comparable hardware resources and power.

中文翻译：

---

图像传感器的混合去雾方法及其硬件实现


在基于图像传感器的户外系统（例如自动驾驶车辆、自动驾驶辅助以及高速公路监控和分析）中，对图像去雾的需求不断增加。这些应用需要专用的硬件解决方案来满足高帧速率和低功耗的限制。之前，已经提出了一些基于现有技术的硬件去雾方法。然而，由于基本假设的失败，它们在恢复的图像中产生了伪影。为了解决这个问题并满足严格的要求，提出了一种基于卷积神经网络（CNN）和暗通道先验（DCP）的数据驱动的图像去雾方法，该方法可以自动学习重要特征并产生更好的结果。所提出的方法是硬件友好的，并且还给出了其硬件实现。该设计采用很少的行缓冲区来存储激活，并消除了对动态随机存取存储器 (DRAM) 等片外存储器的需求。现场可编程门阵列 (FPGA) 和专用集成电路 (AISC) 实现表明，该架构非常适合计算资源受限、内存不足和功耗预算紧张的应用场景。定量分析显示，与最先进的硬件方法相比，在消耗相当的硬件资源和功耗的情况下，标准数据集上的图像质量平均 PSNR 提高了 11% 以上。