Research On Haze Density-Aware Dehazing Methods

When collecting images in severe haze weather,due to the influence of the scattering effect of suspended particles in the environment,the acquired images are seriously degraded.Low-quality degraded hazy images have the characteristics of serious color shift,blurred details,low contrast and poor clarity,which will not only reduce the visual perception,but also affect the accuracy of intelligent systems such as traffic monitoring and automatic driving when inputing hazy images.Therefore,image dehazing technology is particularly important in practical application scenarios.With the rapid development of deep learning technology,various advanced image dehazing algorithms continue to emerge,which effectively improve the image quality after dehazing,but the recovery effect for difficult-to-handle dense hazy images and non-uniform hazy images is not ideal.In addition,low-quality hazy images also have a great impact on subsequent high-level tasks such as object detection,which may cause missed or false detections.Therefore,this paper combines the characteristics of imaging in haze scenes,based on haze density aware,and conducts algorithm research from three aspects: dehazing learning strategy,model architecture design,and high-level detection application task.The main research work is as follows:(1)Aiming at the fact that the difficulty of image dehazing of different densities of haze in real scenes is different,this paper proposes a haze density aware image dehazing algorithm based on curriculum learning and teacher-student learning.By adopting the learning strategy of curriculum learning design from easy to difficult,the teacher-student learning process is constructed,from tinner to heavier hazy images,from synthetic to the real hazy images for image dehazing learning,first using the teacher model to learn the easier tasks,and then through the teacher-student learning,the teacher model helps the student model to learn the more difficult dehazing tasks,and the learning process is better organized in this way.In this paper,a simple and effective method for dividing hard-easy samples of hazy images is proposed,and a knowledge review method based on attention mechanism is established,which effectively improves the learning effect of teacher-student learning.The channel coordinate attention block is proposed to disentangle channel attention and coordinate attention,which significantly improves the computational efficiency of the dehazing model.A large number of experiments show that the proposed algorithm has achieved better results in both quantitative and qualitative aspects.It is worth noting that the curriculum learning strategy proposed in this paper is not limited to specific knowledge distillation methods,which can be combined with different knowledge distillation methods,and has significant performance improvement effects for different deep learning-based dehazing models.(2)Aiming at the static inference mode of traditional deep learning models is difficult to adapt to the variability of hazy images in different scenes and different haze density regions of the same hazy image,this paper proposes an image dehazing algorithm based on dynamic neural network haze density aware.The algorithm introduces dynamic convolution kernels for dynamic convolution calculation,and designs dynamic residual dense block to enhance the expression ability of dehazing models.In addition,the dynamic skip-connected feature fusion component proposed by the algorithm can fuse features at different stages to achieve information complementarity between multi-level features,and then complete efficient dehazing.Experiments show that the proposed algorithm not only achieves a high objective evaluation score,but also reconstructs the dehazing image with good subjective effect,and also has good generalization performance on the cross-domain haze image datasets.(3)The target in the hazy image is obscured by the haze,which leads to the reduced performance of the object detection application task,and the cost of obtaining the annotation information of the target object in the hazy images is high.To solve these problems,this paper proposes a hazy image object detection algorithm based on semi-supervised learning,which uses a large amount of unlabeled hazy images to assist a small amount of labeled hazy images for training by designing a suitable weak-strong data augmentation strategy and semi-supervised teacher-student learning framework,and at the same time performs consistency constraints on the detection results under the two enhancement strategies in an unsupervised manner,so as to encourage the model to learn knowledge from high-quality pseudo-labels to enhance the detection effect of object detectors.Experiments show that the proposed algorithm can effectively improve the performance of the detector in haze scene by using a large amount of unlabeled hazy images.