An Algorithm of Small Object Detection Region Proposal Search Based on GN Splitting

Peiran Zhao; Xinyuan Wu; Xinyu Tang; Xiaohai Shen; Haiyan Xu; Min Li; Xuewu Zhang

doi:10.3788/AOS201838.0915005

Journals >Acta Optica Sinica >Volume 38 >Issue 9 >Page 0915005 > Article

Acta Optica Sinica
Vol. 38, Issue 9, 0915005 (2018)

An Algorithm of Small Object Detection Region Proposal Search Based on GN Splitting

Peiran Zhao^*, Xinyuan Wu, Xinyu Tang, Xiaohai Shen, Haiyan Xu, Min Li, and Xuewu Zhang^*

Author Affiliations

College of Internet of Things Engineering, Hohai University, Changzhou, Jiangsu 213022, China

show less

DOI: 10.3788/AOS201838.0915005 Cite this Article Set citation alerts

Peiran Zhao, Xinyuan Wu, Xinyu Tang, Xiaohai Shen, Haiyan Xu, Min Li, Xuewu Zhang. An Algorithm of Small Object Detection Region Proposal Search Based on GN Splitting[J]. Acta Optica Sinica, 2018, 38(9): 0915005 Copy Citation Text

show less

Fig. 1. Schematic of connected graph

Download full size

Fig. 2. Segment by Quick Shift algorithm. (a) Original image; (b)segmented result

Download full size

Fig. 3. Part of experimental results with Sk=3.5, dmax=13.5. (a) Airplane; (b) sheep; (c) car; (d) buoy

Download full size

Fig. 4. Relationship between time consumption and number of segment regions

Download full size

Algorithm 1: GN_RP
Input: (color) image
Output: Set of small object location boxes B
Obtain initial regions R={r_i}using Quick Shift
Calculate histograms from different color space in r_i
Foreach Neighbouring region pair (r_i.r_j) do
Calculate similarity S(i,j) as the weight w_ij
Generate graph G by neighbor node pair and similarity
Sort w_ij in an ascending order
While w_ij<w_thr do
Delete e_ij from G
Extract set of connected subgraph G'from G_cut
Delete the maximum connected subgraph in G'
Extract object location boxes B from G'

Table 1. Algorithm flow

Types of different parameters	Recall/%	Number of proposals
S_k=5.0, d_max=20.0	70.04	12.06
S_k=3.5, d_max=20.0	71.60	22.17
S_k=5.0, d_max=13.5	81.26	23.38
S_k=3.5, d_max=13.5	84.09	41.14
S_k=2.5, d_max=13.5	78.44	50.49

Table 2. Effects of segmentation parameters on the results

Types of different search	Recall/%	Time/s	Number of proposals
Exhaustion search (9900)	80.55	0.004	9900
SS(S_scale=80, S_min=50)	61.84	0.113	111.86
SS(S_scale=30, S_min=10)	82.89	0.565	700.82
GN_RP(S_k=3.5, d_max=13.5)	84.09	0.068	41.14

Table 3. Comparison of results using different algorithms

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information