Pattern Recognition and Artificial Intelligence

模式识别与人工智能

Home About Journal Editorial Board Instructions Ethics Statement Contact Us 中文

	Judgement and Disposal of Academic Misconduct Article

	Copyright Transfer Agreement

	Proof of Confidentiality

	Requirements for Electronic Version

More....

	Chinese Association of Automation

	National ResearchCenter for Intelligent Computing System

	Institute of Intelligent Machines,Chinese Academy of Sciences

More....

	2022 Vol.35 Issue.10, Published 2022-10-25

	Papers and Reports The Applications of Deep Learning in Image and Vision

The Applications of Deep Learning in Image and Vision

	863	Face Super-Resolution Algorithm Based on Multi-task Adversarial and Antinoise Adversarial Learning
		CHEN Hongyou, CHEN Fan, HE Hongjie, JIANG Tongyu
		The super-resolution(SR) of high magnification single face image is a hard but valuable task. In the face super-resolution(FSR) task, the end-to-end network SR image is fuzzy, and the photoreality and human visual effect are poor. Aiming at the problems, a FSR algorithm based on multi-task adversarial learning(MTAL) and antinoise adversarial learning(ANAL) is proposed. The algorithm is divided into end-to-end network learning and network parameters fine-tuning. To improve the end-to-end learning result, a deep multi-task Laplacian pyramid network(MTLapNet) is designed and integrated with MTAL. The main task is end-to-end learning, while the subtask is the optimization of adversarial learning penalty function. To improve the result of adversarial learning and parameters fine-tuning of the main task network, ANAL is integrated into the optimization process of discriminator of adversarial learning. The experiments show that the proposed algorithm can make the FSR image more photo-realistic and more consistent with human visual habits.
		2022 Vol. 35 (10): 863-880 [Abstract] ( 979 ) [HTML 1KB] [ PDF 2122KB] ( 666 )

	881	Truncated Nuclear Norm Based Unfolding Network for Image Denoising
		ZHANG Lin, YE Hailiang, YANG Bing, CAO Feilong
		In model-driven image denoising, prior regularization terms are required to be constructed in advance, resulting in high computational cost of dealing with optimization models. Data-driven methods possess superior performance and high efficiency due to the flexible architecture and powerful learning capability of neural networks, but their interpretability is insufficient. Therefore, truncated nuclear norm based unfolding network for image denoising is proposed and combined with the model-driven method based on truncated nuclear norm and image denoising in low-rank matrix recovery. Each iteration is regarded as a stage of the unfolding network. Singular value operators are learned with the help of neural networks to solve the problem of expensive computation of singular value decomposition in traditional iterative algorithms. Each of the stages is connected to form an end-to-end trainable unfolding network. The effectiveness of the proposed network is verified by the experiments on multiple datasets of image denoising.
		2022 Vol. 35 (10): 881-892 [Abstract] ( 568 ) [HTML 1KB] [ PDF 2861KB] ( 358 )

	893	Improved Zero-DCE Low-Light Enhancement Algorithm for Night Fatigue Driving Detection
		HUANG Zhenyu, CHEN Yutao, LIN Dingci, HUANG Jie
		Grounded on the existing low-light enhancement algorithm, zero-reference deep curve estimation(Zero-DCE), an improved Zero-DCE low-light enhancement algorithm is proposed to increase the accuracy of night fatigue driving detection. Firstly, the upper and lower sampling structure is introduced to reduce the influence of noise. Secondly, the attention gating mechanism is employed to improve the sensitivity of the network to the face region in the image, and thus the detection rate is increased effectively. Then, an improved kernel selecting module is proposed for the problems arising from noise. Furthermore, standard convolution of Zero-DCE is replaced by the depthwise separable convolution of MobileNet to accelerate the detection. Finally, the driver fatigue state can be judged by the face key point detection network and classification network. The experimental results show that the proposed algorithm improves the accuracies of face detection and eye state recognition rate in a night environment with satisfactory detection results compared with the existing fatigue driving detection algorithms.
		2022 Vol. 35 (10): 893-903 [Abstract] ( 648 ) [HTML 1KB] [ PDF 2666KB] ( 325 )

	904	Unsupervised Cross-Modality Person Re-identification Based on Semantic Pseudo-Label and Dual Feature Memory Banks
		SUN Rui, YU Yiheng, ZHANG Lei, ZHANG Xudong
		The existing supervised visible infrared person re-identification methods require a lot of human resources to manually label the data and they fail to adapt to the generalization of real and changeable application scenes due to the limitation by the labeled data scene. In this paper, an unsupervised cross-modality person re-identification method based on semantic pseudo-label and dual feature memory banks is proposed. Firstly, a pre-training method based on the contrast learning framework is proposed, using the visible image and its generated auxiliary gray image for training. The pre-training method is employed to obtain the semantic feature extraction network that is robust to color changes. Then,semantic pseudo-label is generated by density based spatial clustering of applications with noise (DBSCAN) clustering method. Compared with the existing pseudo-label generation methods, the proposed method makes full use of the structural information between the cross-modality data in the generation process, and thus the modality discrepancy caused by the color change of the cross-modality data is reduced. In addition, an instance-level hard sample feature memory bank and a centroid-level clustering feature memory bank are constructed to make the model more robust to noise pseudo-label by hard sample features and clustering features. Experimental results obtained on two cross-modality datasets, SYSU-MM01 and RegDB, demonstrate the effectiveness of the proposed method.
		2022 Vol. 35 (10): 904-914 [Abstract] ( 465 ) [HTML 1KB] [ PDF 767KB] ( 392 )

	915	Anti-Background Interference Crowd Counting Network Based on Multi-scale Feature Fusion
		YU Ying, LI Jianfei, QIAN Jin, CAI Zhen, ZHU Zhiliang
		With the continuous development of computer vision and artificial intelligence, crowd counting algorithms based on intelligent video analysis have made considerable headway. However, the counting accuracy and robustness are far from satisfactory. Aiming at the problem of multi-scale feature and background interference in crowd counting task, an anti-background interference crowd counting network based on multi-scale feature fusion(AntiNet-MFF) is proposed. Based on the U-Net network architecture, a hierarchical feature split block is integrated into the AntiNet-MFF model, and multi-scale features of the crowd are also extracted with the help of the powerful representation capability of deep learning. To increase the attention of the counting model to the crowd area and reduce the interference of background noise, a background segmentation attention map(B-Seg Attention Map) is generated in the decoding stage. Then, B-Seg attention map is taken as the attention to guide counting model in focusing on the head area to improve the quality of the crowd distribution density map. Experiments on several typical crowd counting datasets show that AntiNet-MFF achieves promising results in terms of accuracy and robustness compared with the existing algorithms.
		2022 Vol. 35 (10): 915-927 [Abstract] ( 655 ) [HTML 1KB] [ PDF 3309KB] ( 331 )

Papers and Reports

	928	Contrastive Learning Based on Bilevel Optimization of Pseudo Siamese Networks
		CHEN Qingyu, JI Fanfan, YUAN Xiaotong
		At present, various designs are applied in contrastive learning algorithms based on pseudo siamese networks to acquire the best student network. However, the performance of teacher network in downstream tasks is ignored. Therefore, an algorithm of contrastive learning based on bilevel optimization of pseudo siamese networks(CLBO)is proposed to acquire the best teacher network by promoting the learning between student and teacher networks. The bilevel optimization strategy includes student network optimization strategy based on nearest neighbor optimization and teacher network optimization strategy based on stochastic gradient descent. The teacher network is regarded as a constraint term through the student network optimization strategy based on nearest neighbor optimization to help the student network learn better from the teacher network. The parameters are calculated by the teacher network optimization strategy based on stochastic gradient descent to update the teacher network. Experiments on 5 datasets show that CLBO performs better than other algorithms in k-NN classification and linear classification tasks. Especially, the advantages of CLBO is obvious when the batch size is smaller.
		2022 Vol. 35 (10): 928-938 [Abstract] ( 536 ) [HTML 1KB] [ PDF 2041KB] ( 341 )

	939	Parallel Incremental Dynamic Attribute Reduction Algorithm Based on Attribute Tree
		QIN Tingzhen, DING Weiping, JU Hengrong, LI Ming, HUANG Jiashuang, CHEN Yuepeng, WANG Haipeng
		Traditional incremental methods mainly focus on the attribute reduction from the perspective of updating approximation. However, while processing large-scale data sets, the methods need to evaluate all attributes and calculate importance repeatedly. Thus, time complexity is increased and efficiency is decreased. To solve the problems, an incremental acceleration strategy for parallelization based on attribute tree is proposed. The key step is to cluster all attributes into multiple attribute trees for parallel dynamic attribute evaluation. Firstly, an appropriate attribute tree is selected for attribute evaluation according to the attribute tree correlation measure to reduce the time complexity. Then, the branch coefficient is added to the stop criterion, and the dynamic increase is conducted with the increase of the branch depth. Consequently, the algorithm can jump out of the cycle automatically after reaching the maximum threshold to avoid the original redundant calculation and improve the efficiency effectively. Based on the above improvements, an incremental dynamic attribute reduction algorithm based on attribute tree is proposed, and a parallel incremental dynamic attribute reduction algorithm based on attribute tree is designed by being combined with Spark parallel mechanism. Finally, experiments on multiple datasets show that the proposed algorithm improves the search efficiency of dynamically variational dataset reduction significantly while maintaining the classification performance, holding a better performance advantage.
		2022 Vol. 35 (10): 939-951 [Abstract] ( 376 ) [HTML 1KB] [ PDF 1104KB] ( 256 )

	952


		2022 Vol. 35 (10): 952-954 [Abstract] ( 203 ) [HTML 1KB] [ PDF 226KB] ( 219 )

模式识别与人工智能

Supervised by
China Association for Science and Technology
Sponsored by
Chinese Association of Automation
NationalResearchCenter for Intelligent Computing System
Institute of Intelligent Machines, Chinese Academy of Sciences
Published by
Science Press

Copyright © 2010 Editorial Office of Pattern Recognition and Artificial Intelligence
Address: No.350 Shushanhu Road, Hefei, Anhui Province, P.R. China Tel: 0551-65591176 Fax:0551-65591176 Email: bjb@iim.ac.cn
Supported by Beijing Magtech 　Email:support@magtech.com.cn