基于卷积神经网络的水面火焰倒影滤波方法

Water surface flame reflection filtering method based on convolutional neural network

为了解决由于水面火焰倒影等区域与火焰特征相似,在水面火焰检测中分割火焰区域时,容易将火焰倒影等区域识别为火焰区域,尤其是当火焰亮度高时,YCbCr空间火焰识别模型可能将较多的火焰亮度高的区域误判为非火焰区域的问题,提出了一种水面火焰识别方法.利用卷积神经网络来判断图片中是否含有火焰,再根据火焰经水面反射、折射后在水中的倒影与火焰区域的亮度差,结合YCbCr空间模型,将火焰区域分割出来.试验结果证明:这种水面火焰识别方法可以判断图像中是否含有火焰图像,能够有效过滤火焰倒影等区域,较为精准地提取出火焰区域;同时可以解决火焰亮度高时,YCbCr空间模型识别火焰区域时存在较大空洞的问题.

The characteristics of flame reflection areas on water surface are similar to those of flame,which are easy to be identified as flame areas during dividing flame areas in water surface flame detection.Especially when the flame brightness is high,the YCbCr space flame recognition model may misjudge areas with high flame brightness as non-flame areas.To solve the problems,a water surface flame recognition method was proposed.The convolutional neural network was used to determine the picture with or without flames.Combined with the YCbCr space model,the flame area was segmented according to the difference between the reflection and refraction of flame in the water and the brightness of flame area after reflection and refraction on the water surface.The test results show that the proposed water surface flame recognition method can determine the image with or without flame images and can effectively filter flame reflections and other areas to extract the flame area more accurately.The problem of YCbCr space model during recognizing the flame area can be solved when the flame brightness is high.