一種平衡準(zhǔn)確性以及高效性的顯著性目標(biāo)檢測(cè)深度卷積網(wǎng)絡(luò)模型

張文明; 姚振飛; 高雅昆; 李海濱

doi:10.11999/JEIT190229

一種平衡準(zhǔn)確性以及高效性的顯著性目標(biāo)檢測(cè)深度卷積網(wǎng)絡(luò)模型

doi: 10.11999/JEIT190229 cstr: 32379.14.JEIT190229

燕山大學(xué)電氣工程學(xué)院秦皇島 066004

基金項(xiàng)目: 河北省自然科學(xué)基金(F2015203212, F2019203195)

詳細(xì)信息

作者簡(jiǎn)介:
張文明：男，1979年生，副教授，研究方向?yàn)楣I(yè)過程控制、機(jī)器視覺

姚振飛：男，1992年生，碩士生，研究方向?yàn)闄C(jī)器視覺與圖像處理

高雅昆：男，1988年生，博士生，研究方向?yàn)闄C(jī)器視覺與圖像處理

李海濱：男，1978年生，教授，研究方向?yàn)楣I(yè)過程控制.、機(jī)器視覺、人工智能

通訊作者:
高雅昆　gaoyakun6@163.com

中圖分類號(hào): TN911.73; TP391.41
計(jì)量
- 文章訪問數(shù): 4203
- HTML全文瀏覽量: 3500
- PDF下載量: 144
- 被引次數(shù): 0
出版歷程
- 收稿日期: 2019-04-08
- 修回日期: 2019-08-30
- 網(wǎng)絡(luò)出版日期: 2020-01-21
- 刊出日期: 2020-06-04

A Deep Convolutional Network for Saliency Object Detection with Balanced Accuracy and High Efficiency

School of Electrical Engineering, Yan Shan University, Qinhuangdao 066004, China

Funds: The Nature Science Foundation of Hebei Province (F2015203212, F2019203195)

摘要

摘要:
當(dāng)前的顯著性目標(biāo)檢測(cè)算法在準(zhǔn)確性和高效性兩方面不能實(shí)現(xiàn)良好的平衡，針對(duì)這一問題，該文提出了一種新的平衡準(zhǔn)確性以及高效性的顯著性目標(biāo)檢測(cè)深度卷積網(wǎng)絡(luò)模型。首先，通過將傳統(tǒng)的卷積替換為可分解卷積，大幅減少計(jì)算量，提高檢測(cè)效率。其次，為了更好地利用不同尺度的特征，采用了稀疏跨層連接結(jié)構(gòu)及多尺度融合結(jié)構(gòu)來提高模型檢測(cè)精度。廣泛的評(píng)價(jià)表明，與現(xiàn)有方法相比，所提的算法在效率和精度上都取得了領(lǐng)先的性能。
- 顯著性檢測(cè) /
- 深度學(xué)習(xí) /
- 分解卷積 /
- 稀疏跨層連接 /
- 多尺度融合
Abstract:
It is difficult for current salient object detection algorithms to reach a good balance performance between accuracy and efficiency. To solve this problem, a deep convolutional network for saliency object detection with balanced accuracy and high efficiency is produced. First, through replacing the traditional convolution with the decomposed convolution, the computational complexity is greatly reduced and the detection efficiency of the model is improved. Second, in order to make better use of the characteristics of different scales, sparse cross-layer connection structure and multi-scale fusion structure are adopted to improve the detection precision. A wide range of evaluations show that compared with the existing methods, the proposed algorithm achieves the leading performance in efficiency and accuracy.
- Saliency detection /
- Deep learning /
- Decomposed convolution /
- Sparse cross-layer connection /
- Multi-scale fusion

HTML全文

圖 1 整體框架圖

下載: 全尺寸圖片幻燈片

圖 2 卷積分解示意圖

下載: 全尺寸圖片幻燈片

圖 3 直連與稀疏跨層連接網(wǎng)絡(luò)結(jié)構(gòu)對(duì)比圖

下載: 全尺寸圖片幻燈片

圖 4 不同連接結(jié)構(gòu)效果對(duì)比圖

下載: 全尺寸圖片幻燈片

圖 5 多尺度融合示意圖

下載: 全尺寸圖片幻燈片

圖 6 不同模型視覺對(duì)比圖

下載: 全尺寸圖片幻燈片

圖 7 5種數(shù)據(jù)集上不同算法P-R曲線圖

下載: 全尺寸圖片幻燈片

表 1 不同卷積結(jié)構(gòu)對(duì)比

結(jié)構(gòu)	參數(shù)量(10⁶)	準(zhǔn)確率(%)	使用時(shí)間(s)
2維卷積	5.16	89.3	0.026
分解卷積	3.75	89.7	0.017

下載: 導(dǎo)出CSV

表 2 不同卷積結(jié)構(gòu)對(duì)比

結(jié)構(gòu)	準(zhǔn)確率(%)	使用時(shí)間(s)
無跨層連接	89.7	0.017
跨層連接	91.7	0.023

下載: 導(dǎo)出CSV

表 3 整體網(wǎng)絡(luò)結(jié)構(gòu)詳表

結(jié)構(gòu)	名稱	類型	輸出尺寸	輸出編號(hào)	結(jié)構(gòu)	名稱	類型	輸出尺寸	輸出編號(hào)
convblock1	reconv$ \times $2		448$ \times $448$ \times $16	1	cross-layer	conv3	rate=12	224$ \times $224$ \times $256	$5" $
cross-layer	conv3	rate=16	448$ \times $448$ \times $32	$1' $	convblock4	maxpool	下采樣
cross-layer	conv3	rate=24	448$ \times $448$ \times $256	$1'' $		reconv$ \times $3		56$ \times $56$ \times $128	6
convblock2	maxpool	下采樣			concat3		融合	56$ \times $56$ \times $256	$(5'+6) $
	reconv$ \times $2		224$ \times $224$ \times $32	2		conv1	降維	56$ \times $56$ \times $128	7
concat1		融合	224$ \times $224$ \times $64	$(1'+2) $	cross-layer	conv3	rate=6	56$ \times $56$ \times $256	$7'' $
	conv1	降維	224$ \times $224$ \times $32	3	convblock5	maxpool	下采樣
cross-layer	conv3	rate=8	224$ \times $224$ \times $64	$3′ $		reconv$ \times $3		28$ \times $28$ \times $256	8
cross-layer	conv3	rate=18	224$ \times $224$ \times $256	$3" $	concat4		融合	28$ \times $28$ \times $1280	$(1''+3''+5''+7''+8) $
convblock3	maxpool	下采樣				conv1	降維	28$ \times $28$ \times $256	9
	reconv$ \times $3		112$ \times $112$ \times $64	4	upblock1	deconv	上采樣
concat2		融合	112$ \times $112$ \times $128	$(3'+4) $		reconv$ \times $3		112$ \times $112$ \times $64
	conv1	降維	112$ \times $112$ \times $64	5	upblock2	deconv	上采樣	448$ \times $448$ \times $2	final
ross-layer	conv3	rate=4	224$ \times $224$ \times $128	$5' $

下載: 導(dǎo)出CSV

表 4 F-measure(F-m)和MAE得分表

算法	MSRA		ECSSD		PASCAL-S		SOD		HKU-IS
算法	F-m	MAE	F-m	MAE	F-m	MAE	F-m	MAE	F-m	MAE
本文方法	0.914	0.045	0.893	0.060	0.814	0.113	0.832	0.119	0.893	0.036
DCL	0.905	0.052	0.890	0.088	0.805	0.125	0.820	0.139	0.885	0.072
ELD	0.904	0.062	0.867	0.080	0.771	0.121	0.760	0.154	0.839	0.074
NLDF	0.911	0.048	0.905	0.063	0.831	0.099	0.810	0.143	0.902	0.048
MST	0.839	0.128	0.653	0.171	0.584	0.236	–	–	–	–
DSR	0.812	0.119	0.737	0.173	0.646	0.204	0.655	0.234	0.735	0.140

下載: 導(dǎo)出CSV

表 5 不同算法處理時(shí)間對(duì)比(s)

模型	本文方法	DCL	ELD	NLDF	MST	DSR
時(shí)間	0.023	1.200	0.300	0.080	0.025	13.580
環(huán)境	GTX1080	GTX1080	GTX1080	Titan X	i7 CPU	i7 CPU
尺寸	448$ \times $448	300$ \times $400	400$ \times $300	300$ \times $400	300$ \times $400	400$ \times $300

下載: 導(dǎo)出CSV

參考文獻(xiàn)(20)

WANG Lijun, LU Huchuan, RUAN Xiang, et al. Deep networks for saliency detection via local estimation and global search[C]. 2015 IEEE Conference on Computer Vision and Pattern Recognition, Boston, USA, 2015: 3183–3192. doi: 10.1109/CVPR.2015.7298938.

LI Guanbin and YU Yizhou. Visual saliency based on multiscale deep features[C]. 2015 IEEE Conference on Computer Vision and Pattern Recognition, Boston, USA, 2015: 5455–5463. doi: 10.1109/CVPR.2015.7299184.

LEE G, TAI Y W, and KIM J. Deep saliency with encoded low level distance map and high level features[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 660–668. doi: 10.1109/CVPR.2016.78.

LIU Nian and HAN Junwei. DHSNet: Deep hierarchical saliency network for salient object detection[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 678–686. doi: 10.1109/CVPR.2016.80.

WANG Linzhao, WANG Lijun, LU Huchuan, et al. Saliency detection with recurrent fully convolutional networks[C]. The 14th European Conference on Computer Vision, Amsterdam, Netherlands, 2016: 825–841. doi: 10.1007/978-3-319-46493-0_50.

ZHANG Xinsheng, GAO Teng, and GAO Dongdong. A new deep spatial transformer convolutional neural network for image saliency detection[J]. Design Automation for Embedded Systems, 2018, 22(3): 243–256. doi: 10.1007/s10617-018-9209-0

ZHANG Jing, ZHANG Tong, DAI Yuchao, et al. Deep unsupervised saliency detection: A multiple noisy labeling perspective[C]. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, 2018: 9029–9038. doi: 10.1109/CVPR.2018.00941.

CAO Feilong, LIU Yuehua, and WANG Dianhui. Efficient saliency detection using convolutional neural networks with feature selection[J]. Information Sciences, 2018, 456: 34–49. doi: 10.1016/j.ins.2018.05.006

ZHU Dandan, DAI Lei, LUO Ye, et al. Multi-scale adversarial feature learning for saliency detection[J]. Symmetry, 2018, 10(10): 457–471. doi: 10.3390/sym10100457

ZENG Yu, ZHUGE Yunzhi, LU Huchuan, et al. Multi-source weak supervision for saliency detection[C]. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Long Beach, USA, 2019: 6067–6076.

SIMONYAN K and ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[J]. 2014, arXiv: 1409.1556.

ALVAREZ J and PETERSSON L. DecomposeMe: Simplifying convNets for end-to-end learning[J]. 2016, arXiv: 1606.05426v1.

LIU Tie, YUAN Zejian, SUN Jian, et al. Learning to detect a salient object[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011, 33(2): 353–367. doi: 10.1109/TPAMI.2010.70

YAN Qiong, XU Li, SHI Jianping, et al. Hierarchical saliency detection[C]. 2013 IEEE Conference on Computer Vision and Pattern Recognition, Portland, USA, 2013: 1155–1162. doi: 10.1109/CVPR.2013.153.

LI Yin, HOU Xiaodi, KOCH C, et al. The secrets of salient object segmentation[C]. 2014 IEEE Conference on Computer Vision and Pattern Recognition, Columbus, USA, 2014: 280–287. doi: 10.1109/CVPR.2014.43.

MOVAHEDI V and ELDER J H. Design and perceptual validation of performance measures for salient object segmentation[C]. 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, San Francisco, USA, 2010: 49–56. doi: 10.1109/CVPRW.2010.5543739.

LI Guanbin and YU Yizhou. Deep contrast learning for salient object detection[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 478–487. doi: 10.1109/CVPR.2016.58.

LUO Zhiming, MISHRA A, ACHKAR A, et al. Non-local deep features for salient object detection[C]. 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, 2017: 6593–6601. doi: 10.1109/CVPR.2017.698.

TU W C, HE Shengfeng, YANG Qingxiong, et al. Real-time salient object detection with a minimum spanning tree[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 2334–2342. doi: 10.1109/CVPR.2016.256.

LI Xiaohui, LU Huchuan, ZHANG Lihe, et al. Saliency detection via dense and sparse reconstruction[C]. 2013 IEEE International Conference on Computer Vision, Sydney, Australia, 2013: 2976–2983. doi: 10.1109/ICCV.2013.370.

相關(guān)文章

施引文獻(xiàn)

資源附件(0)

訪問統(tǒng)計(jì)