改进零参考深度曲线低照度图像增强算法
2023,31(1):209-214
摘要：在低照度条件下拍摄的图像具有对比度低，亮度低，细节缺失等质量缺陷，给图像处理带来困难。提出一种改进零参考深度曲线低照度图像增强算法，通过在空间一致性损失函数中引入与卷积核大小相关参数，统一了不同尺寸图像的增强效果；将颜色不变损失、照明平滑损失函数与输入图像类型关联，使其增强效果的峰值信噪比提高17.75%，对比度提高26.75%；通过使用对称式卷积结构，解决原算法计算量大的问题；通过使用MobileNetV2轻量化网络对零参考深度网络(Zero-DCE)进行了优化，减少网络模型计算复杂度的同时保证模型较好的增强效果。
关键词：深度学习 低照度图像 图像增强 MobileNetV2
LOW-LIGHT IMAGE ENHANCEMENT ALGORITHM BASED ON ZERO-DCE
Abstract：Images taken in low illumination have some quality defects, such as low contrast, low brightness and missing details, which bring difficulties to image processing. An improved zero reference depth curve low illumination image enhancement algorithm is proposed. By introducing the parameters related to the size of convolution kernel into the spatial consistency loss function, the enhancement effect of images with different sizes is unified. The color invariant loss and lighting smoothing loss function are related to the input image type, so that the peak signal-to-noise ratio of the enhancement effect is increased by 17.75% and the contrast is increased by 26.75%. By using symmetric convolution structure, the problem of large computation of the original algorithm is solved. Zero-DCE is optimized by using MobileNetV2 lightweight network, which reduces the computational complexity of the network model and ensures the better enhancement effect of the model.
Key words：Deep learning Low-light images Image enhancement MobileNetV2
收稿日期：2022-10-11
基金项目：江苏省产业前瞻与关键核心技术-碳达峰碳中和科技创新专项资金项目(BE2022044),
