基于图像分割和深度学习的人造板表面缺陷检测.pdfVIP

浙

江

农

林

大

学

学

报，2024，41（1）：176–182

JournalofZhejiangA&FUniversity

doi:

10.11833/j.issn.2095-0756

基于图像分割和深度学习的人造板表面缺陷检测

杨

凡，杨博凯，李荣荣

（南京林业大学

家居与工业设计学院，江苏

南京

210037）

摘要：【目的】针对板式家具零件表面缺陷人工检测过程存在的检测效率低、准确率低、检测结果无法数字化存储等问

题，提出了一种基于图像分割和深度学习算法的饰面人造板表面缺陷的检测方法。【方法】利用工业相机采集人造板图

像，构建缺陷数据集，采用全局阈值和局部动态阈值算法分割表面缺陷与图像截取，通过将ReLU6非线性激活函数替

代ReLU函数，并引入倒残差结构的方法，优化MobileNetv

2深度学习网络，进行缺陷识别与分类。【结果】该方法对

饰面人造板表面崩边和划痕缺陷的检测精确率分别达到了93.1%和97.5%，召回率分别为95.3%和97.6%，单张板件平

均检测用时为163

ms。【结论】本研究提出的方法具有较高精度与稳定性，可解决传统人工检测方法的准确率低、效率

低等问题，为家具板材表面缺陷的自动化检测提供新思路。图6表3参21

关键词：缺陷检测；机器视觉；图像分割；深度学习；板式定制家具

中图分类号：TP29

文献标志码：A

文章编号：2095-0756(2024)01-0176-07

Surface

defect

detection

technology

of

wood-based

panel

based

on

image

segmentation

and

deep

learning

YANG

Fan，YANG

Bokai，LI

Rongrong

（College

of

Furnishings

and

Industrial

Design,

Nanjing

Forestry

University,

Nanjing

210037,

Jiangsu,

China）

Abstract:[Objective]

Aiming

at

the

problems

of

low

detection

efficiency,

low

accuracy

and

digital

storage

of

detection

results

in

the

manual

detection

of

surface

defects

of

panel

furniture

parts,

a

surface

defect

detection

method

of

veneer

wood-based

panel

based

on

image

segmentation

and

deep

learning

algorithm

was

proposed.

[Method]

The

defect

data

set

was

constructed

by

the

artificial

panel

images

collected

by

industrial

cameras.

The

global

threshold

and

local

dynamic

threshold

algorithms

were

used

to

segment

surface

defects

and

image

interceptions.

The

ReLU6

nonlinear

activation

function

was

replaced

by

ReLU

function,

and

the

method

of

reciprocal

residual

structure

was

introduced

to

optimize

the

MobileNetv

2

deep

learning

network,

and

the

defect

identification

and

classification

were

carried

out.

[Result]

The

accuracy

of

the

algorithm

for

the

detection

o