In order to solve the problem that the road image semantic segmentation model has huge parameter number and complex calculation in deep learning, and is not suitable for deployment on mobile terminals for real-time segmentation, a light-weighted symmetric U-shaped encoder-decoder image semantic segmentation network constructed by depthwise separable convolution was introduced, namely MUNet. First, a U-shaped encoder-decoder network was designed; then, the sparse short connection design was added in the convolution blocks; at last, the attention mechanism and Group Normalization (GN) method were introduced to reduce the amount of model parameters and calculation while improving the segmentation accuracy. For the CamVid dataset of road images, after 1 000 rounds of training, the Mean Intersection over Union (MIoU) of the segmentation results of the MUNet was 61.92% when the test image was cropped to a size of 720×720. Experimental results show that compared with the common image semantic segmentation networks such as Pyramid Scene Parsing Network (PSPNet), RefineNet, Global Convolutional Network (GCN) and DeepLabv3+, MUNet has fewer parameters and calculation with better network segmentation performance.

Aiming at the problem that face recognition systems are suspectable to be affected by forgery attacks, a face liveness detection method based on near-infrared and visible binocular vision was proposed. Firstly, the binocular device was used to obtain the face images of near-infrared and visible light synchronously. Then, the facial feature points of two images were extracted, and the binocular relation was used to match the feature points and obtain their depth information, which was used for three-dimensional point cloud reconstruction. Secondly, all facial feature points were divided into four regions, and the average variance of facial feature points in the depth direction within each region was calculated. Thirdly, the key feature points of face were selected. With the nasal tip point as the reference point, the spatial distances between the nasal tip point and the key feature points were calculated. Finally, the feature vectors were constructed by using the depth value variances and spatial distances of facial feature points. And Support Vector Machine (SVM) was used for the judgment of real faces. The experimental results show that the proposed method can detect real faces accurately and resist the attacks of fake faces effectively, achieves the recognition rate of 99.0% in experimental tests, and is superior in accuracy and robustness to the similar algorithm using depth information of facial feature points for detection.

The salient area detected by the existing saliency detection methods has the problems of uneven endoplasm, not clear and accurate boundary. In order to solve the problems, a saliency detection method based on spatial autocorrelation and importance evaluation strategy of graph nodes in complex networks was proposed. Firstly, combined with color information and spatial information, a saliency initial graph under multi-criteria was generated by using the centrality rules of complex network nodes and the spatial autocorrelation indicator coefficient. Then, Dempster-Shafer (D-S) evidence theory was used to fuse multiple initial graphs, and the final salient region results were obtained by adding boundary strength information to a progressively optimized two-stage cellular automaton. The single-step validity verification was performed for each module in the main process of the proposed method on two public image data sets, and the experimental comparisons in visual qualitative index, objective quantitative index and algorithmic efficiency were performed between the proposed method and the other existing saliency detection methods. The experimental results show that, the proposed method is effective in single-step modules, and is superior to other algorithms in terms of the comprehensive results of significant visual effects, Precision-Recall (P-R) curve, F-measure value, Mean Absolute Error (MAE) and algorithm time-consuming, especially to the Background-based maps optimized by Single-layer Cellular Automata (BSCA) algorithm closely related to the proposed method. At the same time, the results of visual contrast experiments also verify that, the proposed method can effectively improve the unsatisfactory results of uneven endoplasm, unclear boundary due to the small difference between salient objects and image background, and the large difference in the internal color of salient objects.

To solve the bad influence on images in foggy days such as decrease of definition and color deviation problem, a fast dehazing method for single image based on dark channel prior was proposed. Firstly, minimum was replaced by gray-scale opening operation to obtain rough dark channel image, and the regions which showed the sudden changes in the foggy image were marked based on the variance so that small window was used to correct the dark channel value in these areas. Next, the rough transmission map was acquired and the guided filter was adopted to refine the transmission. Then, the transmission of sky or other light area was corrected dynamically by using a self-adapting tolerance mechanism. Finally, the final haze-free image was restored from the atmospheric scattering model. The experimental results demonstrate that, compared with other representative dehazing algorithms, the proposed method can achieve a faster processing speed and provide more detailed restored image with good color effect.