International Core Journal of Engineering 2020-26 | Page 153

So far, the basic module of YOLO V3 model has been
introduced. The structure of the model is described below.
Among them, and are the number of grids migrated by
the current grid relative to the upper left corner grid. function
is a logistic function, which normalizes coordinates to 0-1.
and
are normalized values relative to grid
The final
cells. The width and height of anchor box with the greatest
and
coincidence of , ΰ and groundtruth. In practice,
are also normalized to 0-1. In actual program,
and ΰ are
anchor's width, high and featuremap's width and high ratio.
and are normalized values relative to anchor.
The final
416*416*3
DBL
RES1
RES8
DBL
DBL
RES2
RES8
RES4
Upsampling
DBL*5
Darknet-53
concat DBL*5
concat DBL*5
Upsampling
Figure 3 Anchor Candidate Box
The two parts are composed of many small modules. The
modules used in the model are introduced separately below.
DBL
A. Darknetconv2D BN Leaky
DBL module is the basic module of YOLO V3 model, and
its composition is shown in Figure 3.
DBL
=
conv
conv Y1(13*13*15)
DBL conv Y2(26*26*15)
DBL conv Y3(52*52*15)
Leaky relu
BN
Figure 3 DBL module
Among them, conv refers to convolution layer, BN refers
to BN layer, Leaky relu refers to the activation layer whose
activation function is Leaky relu function.
Figure 6 Algorithmic Structural Diagram
B. Res Unit
RU module is the residual structure of Darknet-53
network. It is precisely because of the use of this module that
Darknet becomes deeper and more possible. Its composition
is shown in Figure 4.
RU
=
DBL
DBL
+
Figure 4 RU module
C. RESN Unit
RESN module is a module using residual structure in
Darknet-53 network. Among them, n represents the number,
that is, the positive RESN module uses several layers of RU
module, whose composition is shown in Figure 5.
RESN
=
Zero
padding
DBL
Figure 7 Darknet-53 Network Architecture
As shown in Fig. 6, except that the output y_n is slightly
modified, the other structures are consistent with the structure
of YOLO V3 model, in which upsampling refers to up-
sampling and concat refers to tensor splicing. Unlike the
addition of residual structures, tensor splicing is not a simple
numerical addition, but an expansion of tensor dimension.
RU*n
Figure 5 RESN module
Zero padding refers to zero padding and RU*n refers to
RU modules that repeat n layers.
131