Researches On Image Color Constancy

As an important and effective feature of computer vision,color information has been widely used in many fields of computer vision.However,Image and video collection is affected by illumination in the scene,the reflectivity of the object surface and the response function of the imaging sensor,so it is one of the most unstable features and is easily affected by the illumination in the scene.The purpose of color constancy is to imitate human visual system,which can automatically eliminate the influence of scene illumination on image,restore the true color information of object surface,and provide stable color features for computer vision related research.Based on the theory of computer vision,this dissertation studies the calculation of color constancy under single and multiple illuminations from four aspects.One of the most important applications of color constancy is ”automatic white balance”,and devices using this function are generally embedded devices,such as cameras,network cameras,mobile phones,etc.,which requires fast and stable algorithm and occupies less resources.The first part of this dissertation mainly studies a stable and fast color constancy algorithm which can be applied to low-end embedded devices(Chapter 3).By introducing the dark channel prior de-fogging model into the illumination estimation,a color constancy calculation model based on the dark channel prior model is proposed after a large number of experiments.The algorithm can quickly and accurately extract grey points by designing certain detection criteria.Compared with some existing classical algorithms,it has good performance in subjective vision and objective evaluation,and the computing speed on the low-end(ARM A7 600M)embedded devices is more than 150 frames per second.The development of convolutional neural network has raised the effect of many computer vision related problems to a new level,and has achieved remarkable results in scene recognition,scene segmentation and depth estimation.In order to improve the color constancy in effect,the second,third and fourth parts of this dissertation mainly study the calculation method of color constancy based on convolution neural network for single illumination and multi-illumination.The second part of this dissertation mainly studies the single illumination estimation algorithm based on convolution neural network(the fourth chapter).After analyzing the problems of color constancy algorithm based on convolution neural network,a learning based multi-branch deep probability network(MBDPN)is proposed to estimate the illumination color of the light source in a scene.The network consists of two sub-networks,one is deep multi-branch illumination estimating network(DMBEN),the other is a deep probability computing network(DPN).The multi-branch illumination estimation network can estimate the global illumination and local illuminations through a special multi-branch structure.The other adjoint sub-network DPN separately computes the probabilities results of DMBEN are similar to the ground truth,then determines the better estimation according to the two probabilities under a new criterion.A large number of experimental results show that the network not only improves the effect of illumination estimation,but also outperforms this method in efficiency of 3-4 times.The network model of this structure also provides directions for other work in the future.Human's self-adaptability to natural scene color and prior knowledge of content are important to color constancy.The third part of this dissertation mainly introduces semantic information and multi-scale information into illumination estimation(the fifth chapter).Guided by image semantics,global illumination is estimated from multiple scales by a series of convolution layers and a custom local region illumination convergence layer.The local region illumination convergence layer learns the contribution(weight)of each local region while learning the local illumination,and finally gets the global light at each scale.The illumination obtained from multiple scales is averaged to get the final illumination.Due to the guidance of semantic information,the network is more accurate in learning the contribution of each local region.A large number of experiments shown that the network has greatly improved the effect of illumination estimation.In addition,by using the perceptual distance instead of the traditional loss function to train the optimized network,the convergence speed of the network and the effect of illumination estimation are further improved.Most of the current studies on color constancy assume that only a single illumination exists in the scene.In fact,there is more than one illumination in most scenes,so that the study of color constancy under multi-illumination conditions is more practical.The fourth part of this dissertation mainly studies the color constancy method under complex light sources(the sixth chapter).Under multi-illumination condition,not only the color of illumination but also the location of illumination need to be estimated.In this part,a unique network model with multi-scale supervision and single-scale estimation is designed.In the training process,the learning process is monitored at different scales,while in the illumination estimation process,only the final scale estimation results are calculated.In addition,in the training process,we use a multi-scale computational method with penalty as a loss function to train the network.In the test,the training convergence speed is faster and the illumination estimation effect is better.Through the study of color constancy under single and multi-illumination conditions,several color constancy calculation methods proposed in this dissertation have been verified by a large number of experiments and practical projects.Compared with the existing methods of the same type,both the effect and efficiency have been greatly improved,and has great Practical value.