Research on the Detection of Stellera Chamaejasme Flower Based on
Deep Learning
Jieteng Jiang, Shuo Dong, Chunmei Li
*
and Yihan Ma
Department of Computer Technology and Application, Qinghai University, Xining, Qinghai, 810016, China
Keywords: Deep Learning, YOLOv3-SPP, Faster_Rcnn, SSD, Target Detection, Grassland Degradation.
Abstract: Under the dual impact of global climate change and human activities in recent decades, the grassland
vegetation in the Three-River Source area is seriously degraded, and accurate evaluation of grassland is the
primary condition for ecological protection. Therefore, using intelligent means to evaluate grassland is the
first step in ecological protection. In this paper, the most widely used target detection algorithms Faster-
RCNN, SSD and Yolov3-SPP are used to detect the degradation indicator grass species of Stellera
chamaejasme flower. The experimental results are compared and analyzed, and the characteristics of the three
target detection algorithms and their performance in the detection of degraded indicator grass species are
discussed.
1 INTRODUCTION
In terms of target detection, The Region-based
Convolutional Neural Network (RCNN) (He, Zhang,
Ren, Sun 2016) successfully connects target
detection and deep Convolutional network, and
improves the accuracy of target detection to a new
level. RCNN consists of 3 independent steps:
candidate window generation, feature extraction,
SVM classification and window regression. RCNN
mainly uses the Selective Search method to generate
many candidate windows. Then all the generated
candidate windows are sent to the deep network at
once to extract features. Finally, the SVM classifier
is trained to classify all candidate windows and
window regression. Since RCNN is divided into 3
independent processes, the detection efficiency is
very low. Based on this situation, scholars have
improved RCNN and proposed a scale Spatial
Pyramid Pooling Net (SPPNet) and Fast Region
Based Convolu- tional Neural Network (Fast-RCNN)
(He, Zhang, Ren, Sun 2015). It does not send all the
candidate windows to the network, just send the
image to the deep network once, and then map all the
candidate windows on a certain layer in the network,
which greatly improves the detection speed of the
model. Fast-RCNN (Girshick 2015) uses the
candidate window network (Region Proposal
Network, RPN), and generates candidate windows,
useing the same structure as Fast-RCNN for
classification and window regression. Faster-RCNN
combines target detection into a unified deep network
framework. Region Based Fully Convolutional
Network (RFCN) (Ren, He, Girshick, Sun 2017) is
further improved on this basis. The analysis found
that the network layer after Region of Interest (ROI)
pooling no longer has translation invariance, and the
number of layers after ROI pooling will directly
affect the detection efficiency. Therefore, RFCN
designs a position-sensitive ROI pooling layer, and
directly judges the results after this pooling, which
greatly improves the detection efficiency. YOLO
(You Only Look One) (Dai, Li, He & Sun 2016) and
SSD (Single Shot Multibox Detector) (Redmon,
Divvala, Girshick, Farhadi 2016) are proposed to
improve the detection efficiency of target detection,
and try to make target detection reach the level of
real-time detection. SSD can improve the efficiency
of target detection while maintaining detection
accuracy, which is a win-win algorithm in terms of
detection accuracy and detection efficiency.
Compared with traditional target detection methods,
target detection methods based on deep networks
have obvious advantages in accuracy. First of all, a
neural network is a network structure with self-
learning function that simulates the human brain. The
forward calculation of the deep network can be
regarded as a process of continuously abstracting
objects. The high level of the deep network (near the