孙皓泽 , 常天庆 , 王全东 , 等 . 一种基于分层多尺度卷积特征提取的坦克装甲目标图像检测方法 [J ] . 兵工学报 , 2017 , 38 ( 9 ): 1681 - 1691 . DOI: 10.3969/j.issn.1000-1093.2017.09.003 http://doi.org/10.3969/j.issn.1000-1093.2017.09.003 针对坦克装甲目标的图像检测任务，提出一种基于分层多尺度卷积特征提取的目标检测方法。采用迁移学习的设计思路，在VGG-16网络的基础上针对目标检测任务对网络的结构和参数进行修改和微调，结合建议区域提取网络和目标检测子网络来实现对目标的精确检测。对于建议区域提取网络，在多个不同分辨率的卷积特征图上分层提取多种尺度的建议区域，增强对弱小目标的检测能力；对于目标检测子网络，选用分辨率更高的卷积特征图来提取目标，并额外增加了一个上采样层来提升特征图的分辨率。通过结合多尺度训练、困难负样本挖掘等多种设计和训练方法，所提出的方法在构建的坦克装甲目标数据集上取得了优异的检测效果，目标检测的精度和速度均优于目前主流的检测方法。

SUN H Z , CHANG T Q , WANG Q D , et al . Image detection method for tank and armored targets based on hierarchical multi-scale convolution feature extraction [J ] . Acta Armamentarii , 2017 , 38 ( 9 ): 1681 - 1691 . (in Chinese)

LOWE D G . Distinctive image features from scale-invariant keypoints [J ] . International Journal of Computer Vision , 2004 , 60 ( 2 ): 91 - 110 . DOI: 10.1023/B:VISI.0000029664.99615.94 http://doi.org/10.1023/B:VISI.0000029664.99615.94 http://link.springer.com/10.1023/B:VISI.0000029664.99615.94 http://link.springer.com/10.1023/B:VISI.0000029664.99615.94

BAY H , TUYTELAARS T , GOOL L V . SURF: speeded up robust features [C ] //Proceedings of the European Conference on Computer Vision . Graz, Austria : Springer , 2006 : 404 - 417 .

DALAL N , TRIGGS B . Histograms of oriented gradients for human detection [C ] //Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . San Diego, CA, US : IEEE , 2005 : 886 - 893 .

VAPNIK V N , LERNER A Y . Recognition of patterns with help of generalized portraits [J ] . Avtomatika i Telemekhanika , 1963 , 24 ( 6 ): 774 - 780 . (in Russian)

张珂 , 冯晓晗 , 郭玉荣 , 等 . 图像分类的深度卷积神经网络模型综述 [J ] . 中国图象图形学报 , 2021 , 26 ( 10 ): 2305 - 2325 .

ZHANG K , FENG X H , GUO Y R , et al . Overview of deep convolutional neural networks for image classification [J ] . Journal of Image and Graphics , 2021 , 26 ( 10 ): 2305 - 2325 . (in Chinese)

普运伟 , 刘涛涛 , 郭江 , 等 . 基于卷积神经网络和模糊函数主脊坐标变换的雷达辐射源信号识别 [J ] . 兵工学报 , 2021 , 42 ( 8 ): 1680 - 1689 . DOI: 10.3969/j.issn.1000-1093.2021.08.012 http://doi.org/10.3969/j.issn.1000-1093.2021.08.012 针对人工提取雷达辐射源信号特征耗时长、特征不明显等问题，提出一种基于深度学习卷积神经网络和模糊函数主脊坐标变换的雷达辐射源信号识别方法。该方法通过快速离散分数傅里叶变换提取信号的模糊函数主脊，并将模糊函数主脊极坐标域的二维时频图作为卷积神经网络的输入，实现对不同雷达信号的分选识别。仿真实验结果表明：新方法不仅在信噪比为0 dB以上保持100%的识别率，在-6 dB时识别准确率也稳定在90%以上；相对于传统的雷达信号识别方法和其他深度学习模型识别方法，在识别率和鲁棒性上均有较大提升，具有一定的工程应用价值。

PU Y W , LIU T T , GUO J , et al . Radar emitter signal recognition based on convolutional neural network and coordinate transformation of ambiguity function main ridge [J ] . Acta Armamentarii , 2021 , 42 ( 8 ): 1680 - 1689 . (in Chinese)

呙鹏程 , 吴礼洋 . 融合卷积特征与判别字典学习的低截获概率雷达信号识别 [J ] . 兵工学报 , 2019 , 40 ( 9 ): 1881 - 1889 . DOI: 10.3969/j.issn.1000-1093.2019.09.013 http://doi.org/10.3969/j.issn.1000-1093.2019.09.013 针对低截获雷达信号通常采用人工特征选择，且在低信噪比、样本数量少情况下识别率低的问题，提出一种融合雷达信号时频图像的卷积特征与字典学习识别算法。该算法以表征信号调制方式的时频图像为基础，通过时频变换获得信号的二维时频数据，输入到LeNet-5卷积神经网络中。网络通过美国MNIST数据库手写数据集进行预训练，将预训练后网络中的2～6层网络参数迁移到新的LeNet-5中，取出第6卷积层的数据作为提取的卷积特征。使用判别字典学习方法进行识别。仿真结果表明：通过预训练处理能够加快网络的收敛与优化，有效提取到每类信号的卷积特征；与文献［4］、文献［24］、文献［25］、文献［26］中4种算法相比，利用判别字典学习能够在样本少、低信噪比情况下取得较高的识别率。

GUO P C , WU L Y . LPI radar signal recognition with convolution feature and discrimination dictionary learning [J ] . Acta Armamentarii , 2019 , 40 ( 9 ): 1881 - 1889 . (in Chinese) DOI: 10.3969/j.issn.1000-1093.2019.09.013 http://doi.org/10.3969/j.issn.1000-1093.2019.09.013 The selection of artificial features, low signal-to-noise ratio and small number of samples lead to low recognition rate for low probability of intercepting radar signal. A recognition algorithm with convolution feature and discrimination dictionary learning is proposed. The proposed algorithm is based on the time-frequency image representing a signal modulation type, and a two-dimensional signal is obtained by time-frequency transformation, which is input into LeNet-5. The network is retrained through MNIST data set. The network parameters of 2-6 layers are transferred to a new LeNet-5, and the data from the 6th convolution layer is extracted as convolutional feature. Finally, recognition is ended up by discrimination dictionary learning. Simulated results show that the network goes faster in convergence and optimization through pre-training, and can effectively extract the convolution feature of each kind of signal. Higher recognition rate is obtained through discrimination dictionary learning in the condition of low SNR and small samples compared with other algorithms. Key

刘秋 , 孙晋伟 , 张华 , 等 . 基于卷积神经网络的路面识别及半主动悬架控制 [J ] . 兵工学报 , 2020 , 41 ( 8 ): 1483 - 1493 . DOI: 10.3969/j.issn.1000-1093.2020.08.002 http://doi.org/10.3969/j.issn.1000-1093.2020.08.002 路面对车辆的平顺性、操纵稳定性有直接影响，实时获取路面信息对提升车辆性能具有重要意义。针对传统路面识别方法中难以精确识别多种路面类型的问题，采用卷积神经网络对路面类型进行识别，并根据不同路面输入下悬架系统的输出响应来调整控制器参数，使可控悬架在不同路面下均保持最优性能。建立车辆1/4半主动悬架模型；搭建卷积神经网络基本结构并通过所采集的4种典型城市和非城市路面图像对网络进行训练以及测试；采用遗传算法求取沥青路、弹石路、砂石路、水泥路4种不同路面激励下悬架的最优控制参数；根据路面识别结果及优化结果实现悬架控制参数的自适应调整。仿真结果表明：基于卷积神经网络的路面识别方法能够对多种路面进行准确识别；基于路面识别和遗传算法的半主动悬架控制系统可根据不同路面类型自适应调整悬架参数，有效提升车辆性能。

LIU Q , SUN J W , ZHANG H , et al . Road identification and semi-active suspension control based on convolutional neural network [J ] . Acta Armamentarii , 2020 , 41 ( 8 ): 1483 - 1493 . (in Chinese) DOI: 10.3969/j.issn.1000-1093.2020.08.002 http://doi.org/10.3969/j.issn.1000-1093.2020.08.002 Road has a direct impact on vehicle ride comfort and handling stability, so that the real-time acquisition of road information plays an important role in improving the vehicle performance. The multiple types of road are difficultly identified accurately using traditional road identification methods. The convolutional neural network is used to identify the road type, and then the identified road type is used as the basis for tuning the controller parameters of suspension system in order to make the controllable suspension system maintain the optimal performance under different road surfaces. Firstly, the quarter-vehicle semi-active suspension model is established. Secondly, the basic structure of convolutional neural network is built, and this network is trained and tested based on four typical urban and non-urban road images collected in advance. And then genetic algorithm is used to obtain the optimal control parameters of suspension system under excitations of four different roads, such as asphalt road, sandstone road, pebble road and cement road. Finally, the suspension control parameters are adaptively adjusted according to both the identified and optimized results of road surface. The simulated results show that the road identification method based on convolutional neural network can accurately identify a variety of roads; the semi-active suspension control system based on road identification and genetic algorithm can adaptively adjust the suspension parameters according to different road surfaces, thus improving the vehicle performance effectively.

KRIZHEVSKY A , SUTSKEVER I , HINTON G E . ImageNet classification with deep convolutional neural networks [J ] . Communications of the ACM , 2017 , 60 ( 6 ): 84 - 90 DOI: 10.1145/3065386 http://doi.org/10.1145/3065386 https://dl.acm.org/doi/10.1145/3065386 https://dl.acm.org/doi/10.1145/3065386 We trained a large, deep convolutional neural network to classify the 1.2 million high-resolution images in the ImageNet LSVRC-2010 contest into the 1000 different classes. On the test data, we achieved top-1 and top-5 error rates of 37.5% and 17.0%, respectively, which is considerably better than the previous state-of-the-art. The neural network, which has 60 million parameters and 650,000 neurons, consists of five convolutional layers, some of which are followed by max-pooling layers, and three fully connected layers with a final 1000-way softmax. To make training faster, we used non-saturating neurons and a very efficient GPU implementation of the convolution operation. To reduce overfitting in the fully connected layers we employed a recently developed regularization method called \"dropout\" that proved to be very effective. We also entered a variant of this model in the ILSVRC-2012 competition and achieved a winning top-5 test error rate of 15.3%, compared to 26.2% achieved by the second-best entry.

KOLESNIKOV A , BEYER L , ZHAI X , et al . Big transfer (bit): General visual representation learning [C ] //Proceedings of the European Conference on Computer Vision . Edinburgh, England, UK : Springer , 2020 : 491 - 507 .

穆思奇 , 林进健 , 汪海泉 , 等 . 基于改进YOLOv4的X射线图像违禁品检测算法 [J ] . 兵工学报 , 2021 , 42 ( 12 ): 2675 - 2683 . DOI: 10.3969/j.issn.1000-1093.2021.12.015 http://doi.org/10.3969/j.issn.1000-1093.2021.12.015 为提高安检速度、实现X射线图像中违禁物品的自动检测，提出一种基于改进YOLOv4的X射线图像违禁品检测算法。该算法在单阶段目标检测算法YOLOv4基础上设计一种空洞密集卷积模块。将上采样链路融合后特征输入空洞密集卷积模块中，增强特征表达能力和卷积视野。对融合后特征信息加入注意力机制，用来增强有效特征和抑制无效特征，最终得到表征图像信息的特征图输入检测头部。采用Mosaic数据增强方法训练网络，提升网络的鲁棒性。结果表明：该算法在公开SIXray数据集上的均值平均精度达到80.16%，检测速度为25帧/s；该算法在公开SIXray数据集上多类违禁物品能够取得较高的检测精度，且满足检测的实时性要求。

MU S Q , LIN J J , WANG H Q , et al . An algorithm for detection of prohibited items in X-ray images based on improved YOLOv4 [J ] . Acta Armamentarii , 2021 , 42 ( 12 ): 2675 - 2683 . (in Chinese) DOI: 10.3969/j.issn.1000-1093.2021.12.015 http://doi.org/10.3969/j.issn.1000-1093.2021.12.015 An improved YOLOv4 algorithm for detecting the prohibited items in X-ray images is proposed to increase the speed of security inspection and realize the automatic detection of prohibited items in X-ray images. The proposed algorithm is used to design a dilated dense convolution module based on the one-stage object detection algorithm YOLOv4. The features after the upsampling link fusion are input into the dilated dense convolution module to enhance the feature expression ability and the convolution field of vision. An attention mechanism is added to the fused feature information to enhance effective features and suppress invalid features. Finally，a feature map representing image information is input to detection head. Mosaic data enhancement method is used to train the network to improve the robustness of the network. The results show that the mean average precision (mAP) of the proposed algorithm on the public SIXray data set reaches 80.16%，and the detection speed is 25 frames per second (FPS). The proposed algorithm can achieve high detection accuracy for multiple types of prohibited items on the public SIXray dataset， and meet the real-time requirements of detection.

吴鸿昊 , 王立国 , 石瑶 . 高光谱图像小样本分类的卷积神经网络方法 [J ] . 中国图象图形学报 , 2021 , 26 ( 8 ): 2009 - 2020 .

WU H H , WANG L G , SHI Y . Convolution neural network method for small-sample classification of hyperspectral images [J ] . Journal of Image and Graphics , 2021 , 26 ( 8 ): 2009 - 2020 . (in Chinese)

疏颖 , 毛龙彪 , 陈思 , 等 . 结合自监督学习和生成对抗网络的小样本人脸属性识别 [J ] . 中国图象图形学报 , 2020 , 25 ( 11 ): 2391 - 2403 .

SHU Y , MAO L B , CHEN S , et al . Self-supervised learning and generative adversarial network-based facial attribute recognition with small sample size training [J ] . Journal of Image and Graphics , 2020 , 25 ( 11 ): 2391 - 2403 . (in Chinese)

戴宏 , 郝轩廷 , 盛立杰 , 等 . 面向小样本约束的域适应分类算法 [J ] . 计算机学报 , 2022 , 45 ( 5 ): 935 - 950

DAI H , HAO X T , SHENG L J , et al . Domain adaptation algorithm for few-shot classification task [J ] . Chinese Journal of Computers , 2022 , 45 ( 5 ): 935 - 950 . (in Chinese)

GOODFELLOW I , POUGET-ABADIE J , MIRZA M , et al . Generative adversarial networks [J ] . Communications of the ACM , 2020 , 63 ( 11 ): 139 - 144 . DOI: 10.1145/3422622 http://doi.org/10.1145/3422622 https://dl.acm.org/doi/10.1145/3422622 https://dl.acm.org/doi/10.1145/3422622 \n Generative adversarial networks are a kind of artificial intelligence algorithm designed to solve the\n generative modeling\n problem. The goal of a generative model is to study a collection of training examples and learn the probability distribution that generated them. Generative Adversarial Networks (GANs) are then able to generate more examples from the estimated probability distribution. Generative models based on deep learning are common, but GANs are among the most successful generative models (especially in terms of their ability to generate realistic high-resolution images). GANs have been successfully applied to a wide variety of tasks (mostly in research settings) but continue to present unique challenges and research opportunities because they are based on game theory while most other approaches to generative modeling are based on optimization.\n

黎英 , 宋佩华 . 迁移学习在医学图像分类中的研究进展 [J ] . 中国图象图形学报 , 2022 , 27 ( 3 ): 672 - 686 .

LI Y , SONG P H . Review of transfer learning in medical image classification [J ] . Journal of Image and Graphics , 2022 , 27 ( 3 ): 672 - 686 . (in Chinese)

TAJBAKHSH N , SHIN J Y , GURUDU S R , et al . Convolutional neural networks for medical image analysis: Full training or fine tuning? [J ] . IEEE Transactions on Medical Imaging , 2016 , 35 ( 5 ): 1299 - 1312 . DOI: 10.1109/TMI.2016.2535302 http://doi.org/10.1109/TMI.2016.2535302 Training a deep convolutional neural network (CNN) from scratch is difficult because it requires a large amount of labeled training data and a great deal of expertise to ensure proper convergence. A promising alternative is to fine-tune a CNN that has been pre-trained using, for instance, a large set of labeled natural images. However, the substantial differences between natural and medical images may advise against such knowledge transfer. In this paper, we seek to answer the following central question in the context of medical image analysis: Can the use of pre-trained deep CNNs with sufficient fine-tuning eliminate the need for training a deep CNN from scratch? To address this question, we considered four distinct medical imaging applications in three specialties (radiology, cardiology, and gastroenterology) involving classification, detection, and segmentation from three different imaging modalities, and investigated how the performance of deep CNNs trained from scratch compared with the pre-trained CNNs fine-tuned in a layer-wise manner. Our experiments consistently demonstrated that 1) the use of a pre-trained CNN with adequate fine-tuning outperformed or, in the worst case, performed as well as a CNN trained from scratch; 2) fine-tuned CNNs were more robust to the size of training sets than CNNs trained from scratch; 3) neither shallow tuning nor deep tuning was the optimal choice for a particular application; and 4) our layer-wise fine-tuning scheme could offer a practical way to reach the best performance for the application at hand based on the amount of available data.

ZEBARI R , ABDULAZEEZ A , ZEEBAREE D , et al . A comprehensive review of dimensionality reduction techniques for feature selection and feature extraction [J ] . Journal of Applied Science and Technology Trends , 2020 , 1 ( 2 ): 56 - 70 . DOI: 10.38094/jastt1224 http://doi.org/10.38094/jastt1224 https://jastt.org/index.php/jasttpath/article/view/24 https://jastt.org/index.php/jasttpath/article/view/24 Due to sharp increases in data dimensions, working on every data mining or machine learning (ML) task requires more efficient techniques to get the desired results. Therefore, in recent years, researchers have proposed and developed many methods and techniques to reduce the high dimensions of data and to attain the required accuracy. To ameliorate the accuracy of learning features as well as to decrease the training time dimensionality reduction is used as a pre-processing step, which can eliminate irrelevant data, noise, and redundant features. Dimensionality reduction (DR) has been performed based on two main methods, which are feature selection (FS) and feature extraction (FE). FS is considered an important method because data is generated continuously at an ever-increasing rate; some serious dimensionality problems can be reduced with this method, such as decreasing redundancy effectively, eliminating irrelevant data, and ameliorating result comprehensibility. Moreover, FE transacts with the problem of finding the most distinctive, informative, and decreased set of features to ameliorate the efficiency of both the processing and storage of data. This paper offers a comprehensive approach to FS and FE in the scope of DR. Moreover, the details of each paper, such as used algorithms/approaches, datasets, classifiers, and achieved results are comprehensively analyzed and summarized. Besides, a systematic discussion of all of the reviewed methods to highlight authors' trends, determining the method(s) has been done, which significantly reduced computational time, and selecting the most accurate classifiers. As a result, the different types of both methods have been discussed and analyzed the findings.\n 

GUO Y , SHI H , KUMAR A , et al . Spottune: transfer learning through adaptive fine-tuning [C ] //Proceedings of the IEEE conference on computer vision and pattern recognition . Long Beach, CA, US : IEEE , 2019 : 4805 - 4814 .

NIXON M , AGUADO A . Feature extraction and image processing for computer vision [M ] . Kidlington, UK : Academic press , 2019 .

GUYON I , GUNN S , NIKRAVESH M , et al . Feature extraction: foundations and applications [M ] . Heidelberg, Germany : Springer , 2008 .

BATZELIS E , BLANES J M , TOLEDO F J , et al . Noise-scaled euclidean distance: a metric for maximum likelihood estimation of the PV model parameters [J ] . IEEE Journal of Photovoltaics , 2022 , 12 ( 3 ): 815 - 826 . DOI: 10.1109/JPHOTOV.2022.3159390 http://doi.org/10.1109/JPHOTOV.2022.3159390 https://ieeexplore.ieee.org/document/9747940/ https://ieeexplore.ieee.org/document/9747940/

徐志航 , 陈博 , 张辉 , 等 . 小数据下的音素级别说话人嵌入的语音合成自适应方法 [J ] . 计算机学报 , 2022 , 45 ( 5 ): 1003 - 1017 .

XU Z H , CHEN B , ZHANG H , et al . Speech synthesis adaption method based on phoneme-level speaker embedding under small data [J ] . Chinese Journal of Computers , 2022 , 45 ( 5 ): 1003 - 1017 . (in Chinese)

SCHICK T , SCHÜTZE H . It’s not just size that matters: small language models are also few-shot learners [C ] // Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics:Human Language Technologies . Online : ACL , 2021 : 2339 - 2352 .

LIU Y , REN J H , WU X , et al . Armored vehicle image classification based on transfer learning [C ] //Proceedings of 2021 International Conference on Cyber-Physical Social Intelligence . Beijing, China : IEEE , 2021 : 1 - 5 .

张雪松 , 庄严 , 闫飞 , 等 . 基于迁移学习的类别级物体识别与检测研究与进展 [J ] . 自动化学报 , 2019 , 45 ( 7 ): 1224 - 1243 .

ZHANG X S , ZHUANG Y , YAN F , et al . Status and Development of transfer learning based category-level object recognition and detection [J ] . Acta Automatica Sinica , 2019 , 45 ( 7 ): 1224 - 1243 . (in Chinese)

李策 , 张栋 , 杜少毅 , 等 . 一种迁移学习和可变形卷积深度学习的蝴蝶检测算法 [J ] . 自动化学报 , 2019 , 45 ( 9 ): 1772 - 1782 .

LI C , ZHANG D , DU S Y , et al . A butterfly detection algorithm based on transfer learning and deformable convolution deep learning [J ] . Acta Automatica Sinica , 2019 , 45 ( 9 ): 1772 - 1782 . (in Chinese)

LOGESWARAN L , CHANG M W , LEE K , et al . Zero-shot entity linking by reading entity descriptions [C ] // Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics . Florence, Italy : ACL , 2019 : 3449 - 3460 .

MEHRI S , RAZUMOVSKAIA E , ZHAO T , et al . Pretraining methods for dialog context representation learning [C ] // Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics . Florence, Italy : ACL , 2019 : 3836 - 3845 .

ZHANG W J , WANG J C , LAN F P . Dynamic hand gesture recognition based on short-term sampling neural networks [J ] . IEEE Journal of Automatica Sinica , 2020 , 8 ( 1 ): 110 - 120 .

ZHOU B , KHOSLA A , LAPEDRIZA A , et al . Learning deep features for discriminative localization [C ] //Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . Las Vegas, NV, US : IEEE , 2016 : 2921 - 2929 .

AKIBA T , SANO S , YANASE T , et al . Optuna: a next-generation hyperparameter optimization framework [C ] //Proceedings of the International Conference on Knowledge Discovery and Data Mining . Anchorage, AK, US : ACM , 2019 : 2623 - 2631 .

HE K M , ZHANG X Y , REN S Q , et al . Deep residual learning for image recognition [C ] //Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . Las Vegas, NV, US : IEEE , 2016 : 770 - 778 .

SZEGEDY C , LIU W , JIA Y Q , et al . Going deeper with convolutions [C ] //Proceedings of 2015 IEEE Conference on Computer Vision and Pattern Recognition . Boston, MA, US : IEEE , 2015 : 1 - 9 .