Crowd Density Estimation Method Based on Multi-Feature Information Fusion
Meng Yuebo, Chen Xuanrun, Liu Guanghui, and Xu Shengjun
Crowd density estimation has important application value in the field of intelligent security prevention. A crowd density estimation method with multi-feature information fusion is proposed to address the problems of large difference in viewpoint change of two-dimensional images, loss of feature spatial information, and difficulties in scale feature and crowd feature extraction. The proposed method encodes the multi-view information of images through the attention mechanism-guided perspective of spatial attention (PSA) method to obtain the spatial global contextual information of the feature map and weaken the influence of viewpoint change. Through the multi-scale information aggregation (MSIA) method, the multi-scale asymmetric convolution and the null convolution with different expansion rates are effectively integrated to obtain more comprehensive image scale and feature information. Finally, the spatial information of the high-level feature map and the semantic information of the low-level feature map are complemented by the detailed semantic feature embedding fusion, and the contextual information and scale information complement each other to improve the accuracy and robustness of the model. The experimental validation is carried out using the ShanghaiTech, Mall, and Worldexpo’10 datasets, and the experimental results show that the performance of the proposed method has been improved compared with those of other comparative methods.
  • Oct. 14, 2021
  • Laser & Optoelectronics Progress
  • Vol.58 Issue, 20 2010021 (2021)
  • DOI:10.3788/LOP202158.2010021
Double-Resolution Attention Network for Person Re-Identification
Hu Jiajie, Li Chungeng, An Jubai, and Huang Chao
In person re-identification (ReID) task, some information will be lost in the process of extracting identity-related features, causing the basis for identification become to less and then affects the performance of model. This paper proposes a person ReID method based on double-resolution feature and channel attention mechanism. Firstly, a high-resolution feature branch is added on ResNet, and generate feature vectors corresponding to eight different regions by applying pooling layer on different resolution feature maps. Then a channel attention module is designed based on the situation of feature vectors to enhance the expressive ability of the effective part. Finally, batch normalization is used to coordinate classification loss and measurement loss. In the ablation experiment, the application of each step in the algorithm effectively improves the performance of the model. In the comparative experiments on Market-1501, DUKEMTMC-REID, and CUHK03 datasets, the mean average precision and rank-1 of the proposed algorithm are evidently improved than that of other recent representative algorithms. Experimental results demonstrate that the proposed method can improve the accuracy of person ReID by combining more abundant features.
  • Oct. 13, 2021
  • Laser & Optoelectronics Progress
  • Vol.58 Issue, 20 2010019 (2021)
  • DOI:10.3788/LOP202158.2010019
Iterative Fusion Defogging Algorithm Based on Wavelet Transform
Wei Chunmiao, Xu Yan, and Li Yuan
The image collected outdoors in a foggy environment is prone to low contrast and loss of details. To solve this problem, our study proposes an image fusion method based on multiscale Retinex theory and wavelet transform to restore foggy images. First, the MSR algorithm was used to enhance the collected foggy image. Then, the “db5” wavelet basis was used to merge the brightness component V of the foggy image and the enhanced image, and the saturation component of the foggy image was constrained. Finally, the defogging image was generated. The threshold value was set, and the fog image with relatively high fog concentration was second iteratively fused using wavelet transform to remove residual fog. The experimental results show that the proposed algorithm can effectively restore foggy images with different concentrations. Moreover, the image after fog removal can enhance the details of the dark area, improve the image color, and enrich the image information. Using wavelet fusion preserves more image information so that the image color is rich and natural. Thus, the entire smooth fusion image has a good restoration effect.
  • Oct. 13, 2021
  • Laser & Optoelectronics Progress
  • Vol.58 Issue, 20 2010018 (2021)
  • DOI:10.3788/LOP202158.2010018
Color Zero-Watermarking Algorithm Based on SURF and Halftone Mapping Encryption
Hu Sen, Wu Deyang, Zhong Meiyu, Wang Miaomiao, Zhao Jing, Tang Yong, and Qu Changbo
A color zero-watermarking algorithm based on speeded-up robust features (SURF) to correct geometric distortions and halftone mapping encryption is proposed to address problems associated with poor robustness against geometric attacks in existing digital watermarking algorithms and low efficiency due to excessive embedded color copyright image information. First, the feature points of the carrier image based on SURF algorithm were extracted and feature point information was saved as a key for blind detection. In the copyright verification process, after the feature points of the attacked image were extracted and keys were matched, the geometric correction of the attacked image was implemented based on affine matrix estimated by filtered feature points. Simultaneously, the copyright identification was preprocessed according to the halftone principle. The color image was represented by a three-channel binary matrix with pixel expansion, and the watermark image was encrypted by halftone sub-block mapping according to the encryption rules. In proposed algorithm, the color and structure information of the copyright images were retained while the amount of embedded information was reduced, and the security of the watermark information was increased. The experimental results demonstrate that the proposed zero-watermark algorithm is robust against geometric and nongeometric attacks and the generated zero-watermark information is more secure.
  • Oct. 13, 2021
  • Laser & Optoelectronics Progress
  • Vol.58 Issue, 20 2010017 (2021)
  • DOI:10.3788/LOP202158.2010017
Weak-Light Image Segmentation Based on Intercept Histogram and Otsu Fusion
Zeng Yanyang, Xie Gaosen, and Zhang Jianchun
The line intercept histogram Otsu method has good segmentation performance, but for the weak-light part of the image, it can only divide it into background, and many details of the image are lost. Based on the analysis of the essence of the line intercept histogram algorithm, a directional fuzzy derivative Otsu method is proposed to segment the weak light part of the image. At the same time, in order to enhance algorithm universality, the segmentation results are fused with segmentation results of the traditional Otsu algorithm. Firstly, the directional fuzzy derivatives are used to replace the neighborhood mean of the pixel to achieve a good performance in segmentation of the weak-light part of the image and suppress noise. Then the segmentation results are fused with the Otsu segmentation results to get more accurate threshold segmentation results. Experimental results show that the proposed algorithm can segment the weak-light part of the image more accurately in comparison with other segmentation methods, which provides better noise reduction effects.
  • Oct. 13, 2021
  • Laser & Optoelectronics Progress
  • Vol.58 Issue, 20 2010016 (2021)
  • DOI:10.3788/LOP202158.2010016
Research on Classification Method of Sand and Gravel Aggregate Based on Convolutional Neural Network
Yan Ran, Liao Jideng, Wu Xiaoyong, Xie Changjiang, and Xia Lei
In the current process of detecting commercial sand and gravel aggregates, the manual detection is inefficient, greatly affected by subjective factors, and the detection accuracy is not ideal. This paper proposes a convolutional neural network (CNN) based on the sand and gravel aggregate image classification model (CNN13). This classification model refers to the classic CNN visual geometry group 16 (VGG16) model to improve the network structure and optimize parameters. The CNN13 model uses the TensorFlow deep learning framework to build a 13-layer CNN structure. The experimental dataset includes 5000 digital images, which is collected from the sand and gravel aggregates in the daily production of a commercial concrete manufacturer. The model uses GPU for high-speed calculation during the training process. Compared with the VGG16 model, the CNN13 model has fewer convolutional layers and parameters, lower requirements for GPU memory, faster training speed, higher classification accuracy, and classification accuracy for each level of sand and gravel aggregates is more than 99%.
  • Oct. 13, 2021
  • Laser & Optoelectronics Progress
  • Vol.58 Issue, 20 2010015 (2021)
  • DOI:10.3788/LOP202158.2010015
NSST-Based Perception Fusion Method for Infrared and Visible Images
Li Wei, and Li Zhongmin
To improve the visual perception of fused images, a nonsubsampled shear wave transform (NSST) -based perception fusion method for infrared and visible images is proposed. First, the NSST is used to decompose the source image into high- and low-frequency components. Then, to improve image details, a parameter adaptive pulse coupled neural network is used to fuse high-frequency component images. Second, a Gaussian filter and a bilateral filter are used for multiscale transformation to fuse low-frequency component images, and low-frequency components are decomposed into multiscale texture details and edge features to capture more multiscale infrared spectral features. Finally, the inverse NSST is used to obtain the fused image. Experimental results show that the proposed method can not only improve the detail information of fusion image effectively, but also enhance the ability of infrared feature extraction to fit human visual perception.
  • Oct. 13, 2021
  • Laser & Optoelectronics Progress
  • Vol.58 Issue, 20 2010014 (2021)
  • DOI:10.3788/LOP202158.2010014
Aerial Image Target Detection Algorithm Based on Improved CenterNet
Xu Yanlei, Liang Jiran, Dong Guojun, and Chen Zhuang
In order to improve the accuracy and speed of aerial image target detection, an improved CenterNet aerial image target detection algorithm based on adaptive threshold is proposed. The center point of the target is used as the key point to replace the anchor box for classification and boundary regression, and an adaptive threshold prediction branch is designed to screen and optimize the preprocessing results. At the same time, the encoding-decoding network structure is designed. Through the deformable cavity convolution structure and the convolutional block attention-connection structure based on the attention mechanism, shallow spatial information, and deep semantic information are effectively extracted and fused. In addition, data enhancement is realized by discarding structured information and building new samples with false and missing detection targets, so as to reduce false and missing detection rates. Experiments are performed on the open data set NWPU VHR-10, the results show that compared with CenterNet based on ResNet-50, mean average precision of proposed algorithm increased by 5.17%, and intersection of union of 0.50 and 0.75 are improved by 3.57% and 3.61%, respectively. The detection speed reaches 45 frame·s -1, achieving good detection accuracy and real-time balance.
  • Oct. 13, 2021
  • Laser & Optoelectronics Progress
  • Vol.58 Issue, 20 2010013 (2021)
  • DOI:10.3788/LOP202158.2010013
Attention Fusion Generative Adversarial Network for Single-Image Super-Resolution Reconstruction
Peng Yanfei, Zhang Pingjia, Gao Yi, and Zi Lingling
Deep learning-based single-image super-resolution reconstruction method has been relatively perfect. The reconstructed image has a high objective evaluation value or a good visual effect; however, the image perception effect and objective evaluation value cannot be improved in a balanced manner. To address this problem, this paper proposes a single-image super-resolution reconstruction method based on an attention fusion generative adversarial network. In the proposed method, first, the batch layer that destroys the original image contrast information and affects the quality of image generation in the residual network is removed. Then, the residual block of the attention convolutional neural network, which can effectively perform adaptive feature refinement in the feature map, is constructed. To improve the reconstruction results that lack high-frequency information and texture details under large-scale factors, a pixel-loss function is constructed to replace the mean squared error-loss function with a more robust Charbonnier loss function, and a total variation regular term is used to smooth the training results. The experimental results show that compared with other methods on the Set5, Set14, Urban100, and BSDS100 test sets under 4× magnification factor, the average peak signal-to-noise ratio and average structure similarity increased by 2.88 dB and 0.078, respectively. The experimental data and renderings demonstrate that the proposed method is subjectively rich in details, objectively has a high peak signal-to-noise ratio and structural similarity value, and achieves a balanced improvement of visual effects and objective evaluation index values.
  • Oct. 13, 2021
  • Laser & Optoelectronics Progress
  • Vol.58 Issue, 20 2010012 (2021)
  • DOI:10.3788/LOP202158.2010012
Research of Dehazing Algorithm Based on Dark Channel Prior
Pu Hengfei, Li Zhen, and Li Liangrong
Aiming at the problems of block effect and high algorithm complexity in the dark channel prior dehazing algorithm, we propose an improved dehazing algorithm based on the dark channel prior. First, row transmittance is obtained through the dark channel prior dehazing algorithm, and then the haze parameters are adjusted adaptively by the peak signal-to-noise ratio to obtain the optimized transmittance. Then, the above results are trained as the input vector and the target vector of the multilayer perceptron to establish the mapping between the row transmittance and the optimized transmittance and obtain the optimized transmittance. Finally, the image is restored by the atmosphere light value to obtain the haze-free image. Experimental results show that the algorithm can effectively improve the block effect, improve the restoration efficiency, and improve the clarity of image details to a certain extent.
  • Oct. 13, 2021
  • Laser & Optoelectronics Progress
  • Vol.58 Issue, 20 2010011 (2021)
  • DOI:10.3788/LOP202158.2010011