ing at the existing remote sensing aircraft image detection methods based on deep learning, which require a large number of tagged data sets and a long training time, we propose a semi-supervised learning method based on generative adversarial networks (GANs). Two granularity deep-learning generative adversarial networks are used to get the edge feature and deep semantic feature information. By combining these two discriminator networks of the GANs, we design the object detection model. The experiment shows that the proposed method has a faster training speed and less labeled dataset is needed during the training process.

ing at the problems of the halo phenomenon and inaccurate selection of atmospheric light values in dark channel prior algorithm, an image dehazing method based on dark channel compensation and improvement of atmospheric light value is proposed in this paper. In order to weaken the halo effect at the edge of the image scene, a solution based on the dark channel compensation model is proposed first, the halo region is identified by the weighted channel difference method, and then the dark channel values of this region are modified by corrosion, fusion, and other treatment. It is linearly fused with the original dark channel images to compensate the dark channel. For the problem of inaccurate selection of atmospheric light value, the quadtree segmentation method is improved, with the strategy of adjacent region comparison added. Hence, the proposed method can obtain more accurate atmospheric light values, leading to more clear and natural restored images with more details. Finally, the haze-free image is restored by means of the atmospheric scattering model and the optimized transmittance. The experimental results show that the proposed method can effectively remove the halo effect and obtain the atmospheric light value accurately.

ing at solving the problems of color and contrast distortions in traditional dehazing algorithms, we propose an image dehazing algorithm based on multi-scale fusion and adversarial training. The multi-scale feature extraction block is used to extract haze-relevant features from different scales, and the residual-and-densely-connected block is used to realize the interaction of image features and avoid gradient vanishing. Because the algorithm is not based on the atmospheric scattering model and directly fuses the shallow and deep features of the image in the multi-scale manner, so it overcomes the inaccuracy of physical model. The dehazing network is trained via the generative adversarial mechanism, the generator uses the multi-scale feature extraction block and the residual-and-densely-connected block to estimate the haze-free image, and the discriminator consisting of two sub-networks with different receptive fields carries out the adversarial training. Comparison experiments on the RESIDE (Realistic single image dehazing) dataset show that the dehazed images generated by the proposed algorithm are more visually pleasant than those by other algorithms in terms of full-reference and no-reference visual quality indicators.

ing at the industrial images with uneven illumination interference, a local threshold segmentation method based on multi-direction grayscale wave for image is proposed in this paper. First, the image is pretreated with mean filtering. Then, one-dimensional grayscale wave curves are extracted in four-direction of horizontal, vertical, left diagonal and right diagonal. Meanwhile, one-dimension local threshold segmentation is carried out for wave peaks and troughs which satisfy the threshold condition of wave amplitude on each curve. Finally, an and operation is processed for segmented sub-images, the final segmented image is obtained. The experimental results show that this method can effectively improve the segmentation accuracy of images with uneven illumination. Compared with two-dimensional Otsu, two-dimensional Tsallis and Niblack methods, the segmentation effect of the proposed method is significantly improved.

ing at the characteristics of high dimensionality of the hyperspectral image data, nonlinearity of the feature and difficulty of obtaining the tag data, combined with the stack sparse automatic coding network, we propose a two-level classification algorithm based on nonlocal mode feature fusion. Compared with the traditional stack sparse automatic coding network, the spectral angle matching algorithm stacks the spectral information found most similar to the classified pixel to form new spectral information, and puts it into the SoftMax classifier for first-level classification. The pixels satisfying the condition are added to the training data set for classification training of the stack sparse coding network. Finally, the classification algorithm is modified according to the spatial neighborhood information to make the classification result more smooth. Compared with other classification algorithms, it is found that the improved classification algorithm has higher accuracy and can effectively improve the classification effect of hyperspectral image.

ing at the problem that nonuniform cloud is difficult to remove effectively by using single algorithms for high resolution remote sensing satellite images, an optimization algorithm based on image segmentation and improved dark channel prior method is proposed. The original cloud image is segmented into a dense fog area and a thinner fog area by the image segmentation technique. The dense fog area adopts weighted multiscale Retinex algorithm to realize local enhancement and remove the fog. The thinner fog area adopts the improved dark color method, transforming the dark color image defogging model from RGB color space to HSI color space, extracting the luminance component, and obtaining accurate atmospheric optical values. The atmospheric transmittance is optimized by the tolerance mechanism, and the defogged image is obtained by enhancement of the automatic gradation method. Experimental results show that the proposed algorithm can restore image details and recover image color and clarity effectively.

ing at the problem of high complexity and low processing speed of heterologous image registration algorithms, a fast registration algorithm of visible light and synthetic aperture radar (SAR) images is proposed. In the image preprocessing stage, the redundant information in visible light and SAR images is removed, and two different types of images are filtered respectively with Gauss low-pass filter and non-local mean filter (NLM) algorithms. Then, the multi-scale Harris method is used to detect and extract feature points, and the gradient position orientation histogram (GLOH) method is used to construct descriptors of feature points. Finally, the feature points in the original image are reconstructed based on the feedback mechanism, and the actual position of the feature points to be matched in the original image is got, so as to complete the reconstruction and matching of the feature points in the original image. The experimental results show that compared with scale invariant feature transform-modification (SIFT-M) method, this algorithm significantly reduces the running time while maintaining the average registration accuracy of more than 80%, and has important application value.