The visibility and contrast of images with haze are often relatively low.The higher the concentration of haze in the image,the more fuzzy the image will be,which has a bad effect on various image processing tasks such as target detection,semantic segmentation and image classification.Therefore,as an effective image preprocessing step,image dehazing has been widely concerned by computer vision and graphics researchers.In order to solve the problem of image dehazing,this thesis proposes an improved single image dehazing method based on dark channel prior dehazing method and an improved cycle generation adversarial network model for image dehazing.The main contents of this thesis are as follows:(1)Starting from the study of the Dark Channel Prior dehazing method,this thesis finds that the calculation method of atmospheric light value is easy to be affected by bright objects in the image,the detailed transmittance by the soft-cut method will lose the detailed information of the image and take a long time.and the sky area is prone to image over-enhancement.For these problems,this thesis has optimized: According to the proposed sky region segmentation method,the sky region and non-sky region in the input image are accurately divided,and then the adaptive atmospheric light threshold is calculated separately.The final atmospheric light value is obtained by averaging pixels in the sky region that are larger than the threshold.The obtained value is not affected by the bright objects in the image and is closer to the real value.Finally,guided filtering is used to refine the transmittance instead of soft-cut method to refine the transmittance and compensate for the transmittance of the sky area.After refining,the transmittance map is clearer and retains more image details,which effectively reduces the time spent in refining the transmittance map.The final experiment proves that the optimized method has good performance in dehazing,the image after dehazing is more real and natural,the visual effect is outstanding,and higher objective evaluation criteria peak signal-to-noise ratio and structural similarity index measure scores are obtained,and the running speed is increased by nearly 30%.(2)At present,the dehazing model based on deep learning generally relies on paired images with and without haze as training data sets,however,the collection of paired data sets is difficult.Therefore,almost all images with haze in paired data sets are generated after artificial dehazing.In this way,haze features learned by the model are often different from those in reality.As a result,the dehazing effect of the model on natural hazed images is not ideal.In order to improve the dehazing performance of the model,the cyclic generation adversarial network is improved and applied in the field of image dehazing: Firstly,the discriminator in the network is replaced by a Markov discriminator,the layers of the generator network are deepened,meanwhile,the pooling layer in the network is replaced by a convolution layer with step size of 2.These improvements improves the feature extraction ability of the generator and alleviates the problem of gradient disappearance.Finally,a new combinatorial loss function is proposed for image dehazing to improve the effect of dehazing.Relevant ablation experiments prove that the modification of the improved model in the network architecture and loss function is successful.In the experimental comparison with other dehazing algorithms,the model not only achieves the best visual effect in the dehazing task,but also gets the highest objective evaluation criteria PSNR and SSIM. |