~目标检测在智能交通、智慧城市、智能安防和遥感探测等诸多领域得到广泛的应用。但是，当需要检测的目标密集出现在同一区域时，如在拥堵的道路上检测大量的车辆时，现有目标检测算法的检测效率不佳，容易出现错识、漏识等现象，导致检测率不理想，往往不能准确的检测并统计出需要检测的目标数量。因此本文针对在同一区域密集出现的相似目标，在现有的目标检测技术的基础上，通过使用网络特征热力图优化VGG-16卷积神经网络进行特征提取，获得同一区域内密集出现的相似目标的结构特征和空间特征，有效检测密集相似目标，最后结合边缘检测技术获取检测目标区域，达到检测相似目标的理想效果。主要工作如下：
传统的卷积神经网络的目标检测算法是一个典型的黑箱问题，网络在训练中只有输入和输出，人工无法干预网络的训练过程，因此针对这一问题，本文将网络特征热力图应用在卷积神经网络训练的每一步，指导网络训练。网络特征热力图可以反映神经网络训练中网络特征的富集情况，最终通过提取特征分类候选区完成目标检测。实验结果表明，在密集相似目标检测领域，通过网络特征热力图优化的卷积神经网络相对于传统的神经网络，在训练效率和目标检测率上都有提升。
针对特征提取模块网络模型单一、目标特征提取不充分等问题，本文改进了一种基于多尺度卷积的密集目标识别网络架构，通过建立具有不同卷积核的多尺度卷积网络进行特征提取，然后将提取到的多尺度特征输入到优化后的目标识别网络中进行目标识别，最后通过分类优化后的候选区，完成相似目标检测。通过实验证明，本文优化的方法在密集相似目标检测领域取得了进一步提升，本文优化后的方法对比于传统的目标检测方法精确率提高了2.44个百分点，同时漏选率降低1.82个百分点，与第三章相比精确率提高了0.24个百分点，同时漏选率降低0.45个百分点。

 Target detection is widely used in intelligent transportation, intelligent city, intelligent security and remote sensing. However, when the targets to be detected are concentrated in the same area, for example, when a large number of vehicles are detected on the congested road, the detection efficiency of the existing target detection algorithm is not good, prone to misidentification, missing identification and other phenomena, resulting in the unsatisfactory detection rate, often unable to accurately detect and count the number of targets to be detected. So this paper in the same area intensive appear similar goals, on the basis of the existing target detection technology, through the use of the network characteristics of hot tries to optimize VGG - 16 convolution neural network for feature extraction, obtain dense appear similar goals within the same area of structure characteristics and spatial characteristics, effective detection of dense similar goals, Finally, the edge detection technology is used to obtain the detection target region to achieve the ideal effect of similar target detection. The main work is as follows:
The target detection algorithm of traditional convolutional neural network is a typical black box problem. The network only has input and output in the training, and manual cannot intervene in the training process of the network. Therefore, to solve this problem, this paper applies the network characteristic thermogram in each step of the convolutional neural network training to guide the network training. Network characteristic thermal map can reflect the enrichment of network features in neural network training, and finally achieve target detection by extracting feature classification candidate areas. Experimental results show that compared with traditional neural networks, the convolutional neural network optimized by network characteristic thermal map has improved training efficiency and target detection rate in the field of dense similar target detection.
In view of the feature extraction module network model of the single, such problems as inadequate target feature extraction, this paper improves a dense target recognition based on multi-scale convolution network architecture, through the establishment of multi-scale convolution networks with different convolution kernel feature extraction, and then extracted to multi-scale characteristics of the input to the optimized target identification in the network is used to identify the target, Finally, similar target detection is completed through the candidate area after classification optimization. Experiments show that the optimized method in this paper has further improved in the field of dense similar target detection. Compared with the traditional target detection method, the optimized method in this paper improves the accuracy by 2.44 percentage points, while the missed selection rate is reduced by 1.82 percentage points, and the accuracy is improved by 0.24 percentage points compared with the third chapter. At the same time, the missed separation rate decreased by 0.45 percentage points.
 

[1] 曹金梦,倪蓉蓉,杨彪.基于多尺度多任务卷积神经网络的人群计数[J].计算机应用,2019,39(01):199-204.

[2] 胡聪,屈瑾瑾,许川佩,朱爱军.基于自适应池化的神经网络的服装图像识别[J].计算机应用,2018,38(08):2211-2217.

[3] 卫星,乐越,韩江洪,陆阳.基于长短期记忆的车辆行为动态识别网络[J].计算机应用,2019,39(07):1894-1898.

[4] 王理顺,钟勇,李振东,贺宜龙.基于深度学习的织物缺陷在线检测算法[J].计算机应用,2019,39(07):2125-2128.

[5] 李旭冬, 叶茂, 李涛. 基于卷积神经网络的目标检测研究综述[J]. 计算机应用研究, 2017, 34(010):2881-2886.

[6] 陈乐,童莹,陈瑞,曹雪虹.端到端增强特征神经网络的视频表情识别[J].重庆理工大学学报(自然科学),2019,33(09):125-131.

[7] 刘雨桐,李志清,杨晓玲. 改进卷积神经网络在遥感图像分类中的应用[J].计算机应用,2018,38(4):949-954.

[8] 罗元,云明静,王艺,赵立明.基于人眼信息特征的人体疲劳检测[J].计算机应用,2019,39(07):2098-2102.

[9] 熊咏平,丁胜,邓春华, 等.基于深度学习的复杂气象条件下海上船只检测[J].计算机应用,2018,38(12):3631-3637. DOI:10.11772/j.issn.1001-9081.2018040933.

[10] 唐浩漾,孙梓巍,王婧,钱萌.基于VGG-19混合迁移学习模型的服饰图片识别[J].西安邮电大学学报,2018,23(06):87-93.

[11] 赵蓬辉,孟春宁,常胜江.基于改进VGG网络的单阶段船舶检测算法[J].光电子·激光,2019,30(07):719-730.

[12] 刘国辉,张伟伟,吴训成,宋晓琳,许莎,温培刚.基于VGG-M网络模型的前方车辆跟踪[J].汽车工程,2019,41(01):57-63.

[13] 张建华,孔繁涛,吴建寨,翟治芬,韩书庆,曹姗姗.基于改进VGG卷积神经网络的棉花病害识别模型[J].中国农业大学学报,2018,23(11):161-171.

[14] 陈毅,黄妙华,胡胜.基于改进VGG卷积神经网络的前方车辆目标检测[J].数字制造科学,2018,16(04):282-287.

[15] Lu Z, Jiang X, Kot C C. Deep Coupled ResNet for Low-Resolution Face Recognition[J]. IEEE Signal Processing Letters, 2018:526-530.

[16] Xue F, Ji H, Zhang W. Mutual information guided 3D ResNet for self-supervised video representation learning. 2020.

[17] Sun W, Yao B, Chen B, et al. Noncontact Surface Roughness Estimation Using 2D Complex Wavelet Enhanced ResNet for Intelligent Evaluation of Milled Metal Surface Quality[J]. Applied Sciences, 2018, 8(3):381.

[18] Chen W, Huang H, Peng S, et al. YOLO-face: a real-time face detector[J]. The Visual Computer, 2020(9).

[19] Hassner T, Liu C. SIFT Flow: Dense Correspondence across Scenes and its Applications[J]. Springer International Publishing, 2016, 10.1007/978-3-319-23048-1(Chapter 2):15-49.

[20] Mizuno K, Terachi Y, Takagi K, et al. Architectural Study of HOG Feature Extraction Processor for Real-Time Object Detection[C]// Signal Processing Systems (SiPS), 2012 IEEE Workshop on. IEEE, 2013.

[21] Tareen S A K, Saleem Z. A comparative analysis of SIFT, SURF, KAZE, AKAZE, ORB, and BRISK[C]// 2018 International Conference on Computing, Mathematics and Engineering Technologies (iCoMET). IEEE, 2018.

[22] Grave E, Joulin A, Cissé, Moustapha, et al. Efficient softmax approximation for GPUs[J]. 2016.

[23] Sun X, Wu P, Hoi S. Face Detection using Deep Learning: An Improved Faster RCNN Approach[J]. Neurocomputing, 2018, 299(JUL.19):42-50.

[24] Redmon J, Farhadi A. YOLOv3: An Incremental Improvement[J]. arXiv e-prints, 2018.

[25] Lin T Y, Dollar P, Girshick R, et al. Feature Pyramid Networks for Object Detection[C]// 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society, 2017.

[26] 赵娜. 基于卷积神经网络的目标检测算法综述[J]. 新一代信息技术, 2019, 002(024):50-55.

[27] 完颜幸幸,傅钦翠,吉鑫.迁移学习用于牵引变电所视频多目标识别研究[J].计算机工程与应用,2019,55(24):196-201.

[28] 高宗,李少波,陈济楠,李政杰.基于YOLO网络的行人检测方法[J].计算机工程,2018,44(05):215-219+226.

[29] 盛恒,黄铭,杨晶晶.基于Faster R-CNN和IoU优化的实验室人数统计与管理系统[J].计算机应用,2019,39(06):1669-1674.

[30] 周晓彦, 王珂, 李凌燕. 基于深度学习的目标检测算法综述[J]. 电子测量技术, 2017(11).

[31] 张汇, 杜煜, 宁淑荣,等. 基于Faster RCNN的行人检测方法[J]. 传感器与微系统, 2019, 038(002):147-149,153.

[32] 张亚倩. 卷积神经网络研究综述[J]. 信息通信, 2018, No.191(11):32-34.

[33] Ren S, He K, Girshick R, et al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6):1137-1149.

[34] 白云汉. 基于SLAM算法和深度神经网络的语义地图构建研究[J]. 计算机应用与软件, 2018(1):183-190.

[35] Li L, Si Y, Jia Z. Medical Image Enhancement Based on CLAHE and Unsharp Masking in NSCT Domain[J]. Journal of Medical Imaging and Health Informatics, 2018.

[36] Fu Q, Celenk M, Wu A. An improved algorithm based on CLAHE for ultrasonic well logging image enhancement[J]. Cluster Computing, 2018.

[37] 章慧, 龚声蓉. 基于改进的Sobel算子最大熵图像分割研究[J]. 计算机科学, 2011(12):278-280.

[38] 段红燕,邵豪,张淑珍,张晓宇,王小宏.一种基于Canny算子的图像边缘检测改进算法[J].上海交通大学学报,2016,50(12):1861-1865.

[39] 王嘉俊, 段先华. 改进Canny算子在水面目标边缘检测中的研究[J]. 计算机时代, 2020, 000(001):35-38.

[40] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, Dhruv Batra. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization[J]. Springer US,2020,128(2).

[41] Chattopadhyay A, Sarkar A, Howlader P, et al. Grad-CAM++: Generalized Gradient-based Visual Explanations for Deep Convolutional Networks[J]. 2017.

[42] Simonyan K, Zisserman A. Very Deep Convolutional Networks for Large-Scale Image Recognition[J]. Computer Science, 2014.

[43] 姚明海,杨圳.基于轻量级卷积神经网络的实时缺陷检测方法研究[J].计算机测量与控制,2019,27(06):22-25+40.

[44] Liu N, You J. Research on Vehicle Detection Based on Visual Convolution Network Optimization.[J] Verification and Evaluation of Computer and Communication Systems. VECoS 2020. Lecture Notes in Computer Science, vol 12519. Springer, Cham; 237-243.

[45] Szegedy C, Vanhoucke V, Ioffe S, et al. Rethinking the Inception Architecture for Computer Vision[J]. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016:2818-2826.