Vision-Based Defect Classiﬁcation and Weight Estimation of Rice Kernels Xiang Wang1 Kai Wang1 Xiaohong Li1 and Shiguo Lian1

2025-05-06 0 0 2.21MB 9 页 10玖币

侵权投诉

Vision-Based Defect Classiﬁcation and Weight Estimation of Rice

Kernels

Xiang Wang1, Kai Wang1, Xiaohong Li1, and Shiguo Lian1

Abstract— Rice is one of the main staple food in many areas

of the world. The quality estimation of rice kernels are crucial

in terms of both food safety and socio-economic impact. This

was usually carried out by quality inspectors in the past, which

may result in both objective and subjective inaccuracies. In this

paper, we present an automatic visual quality estimation system

of rice kernels, to classify the sampled rice kernels according

to their types of ﬂaws, and evaluate their quality via the weight

ratios of the perspective kernel types. To compensate for the

imbalance of different kernel numbers and classify kernels with

multiple ﬂaws accurately, we propose a multi-stage workﬂow

which is able to locate the kernels in the captured image and

classify their properties. We deﬁne a novel metric to measure

the relative weight of each kernel in the image from its area,

such that the relative weight of each type of kernels with regard

to the all samples can be computed and used as the basis for

rice quality estimation. Various experiments are carried out

to show that our system is able to output precise results in a

contactless way and replace tedious and error-prone manual

works.

Index Terms— Rice defect classiﬁcation, rice quality estima-

tion, rice weight estimation, vision-based method.

I. INTRODUCTION

Rice is the staple food for people in many parts of the

world, and its quality is related to food health as well as the

economic interests of agricultural dealers, who make offer

to the farmers according to the quality of the collected rice

kernels. Manual rice quality estimation is quite tedious, as

it requires experienced inspectors to identify and pick up

the kernels with various defects one by one and weigh them

carefully. The precision of the result is subject to the skill

and conscientiousness of the inspectors.

With the development of computer vision technique in

recent decades, lots of attempts have been made to classify

rice types and defects automatically [1]. A wide range of

computer vision algorithms, from traditional geometric [2] to

deep-learning based methods [3] have been utilized. However,

some inherent problems in precise rice quality analysis still

exist. For example, rice kernels have various properties

divided according to different features, and there are even

some kernels containing more than one kinds of defects. In the

past, the classiﬁcation work often relies on the back and forth

manual checks of the inspectors. Meanwhile, the estimation

of weight ratios of each type of kernel, which is used as the

key metric for rice quality and was usually measured via

high-precision digital steelyards, is not easy to perform only

All authors are with AI Innovation and Application Center,

China Unicom.

wangx386, wangk115, lixh585,

liansg@chinaunicom.cn

through its appearance. The reason behind is that although

state-of-the-art artiﬁcial intelligence has been able to achieve

remarkable results in various ﬁelds, manual rules and post-

processing works are still indispensable in practice to make it

applicable enough to replace skilled engineers and workers.

Fig. 1: Rice quality estimation results with the proposed

system. The type of each rice kernel is labeled on its bounding

box, and the weight ratios are shown in the upper right area.

PC, MC, YC, SP, BR and SO represent Partial-Chalky, Mass-

Chalky, Yellow-Colored, Spotted, Broken and Sound kernels

respectively.

To solve these problems which existing vision-based

methods are not able to handle, we propose a novel approach

for automatic quality estimation of rice kernels. We designed a

hardware system to obtain high-quality images of rice kernels

with little lighting interference and background noise. To

detect and classify the kernels with various types of defects,

we employed a multi-stage classiﬁcation approach to perform

multi-classiﬁcation of rice kernel ﬂaws, such that a single

kernel with dual defects can be detected and the classiﬁcation

accuracy for different defects is improved. Then we use

a vision-based method to estimate the relative weights of

kernels with each type of defect, by measuring its density via

a novel weight-per-pixel metric and multiplying it with the

segmented areas. An example of our result is shown in Fig. 1.

With such workﬂows, the quality of the rice kernels can be

automatically analyzed and presented to the inspectors, and

tedious manual works can be replaced.

The contributions of this paper include:

•

The introduction of a multi-stage method for detection

and classiﬁcation of rice kernels with various defects;

arXiv:2210.02665v1 [cs.CV] 6 Oct 2022

•

The proposal of a weight-per-pixel metric for the

estimation of weight ratios of each type of kernels;

•

The design of a complete system which integrates

these approaches to produce high-precision rice quality

evaluation results automatically.

The rest of the paper is organized as follows: Section II

reviews existing methods for rice classiﬁcation and defect

detection. Section III introduces the details of the proposed

system and methods. Experimental evaluation is presented in

Section IV, and Section Vgives the conclusions.

II. RELATED WORK

During the last decades, various approaches for automatic

rice classiﬁcation and quality analysis have been proposed.

Rice classiﬁcation:

Most work adopts the following route:

extracting features with image processing algorithms and then

classifying the rice based on these features. Kuo et al. [4]

classiﬁed 30 varieties rice grains using image processing and

sparse-representation-based classiﬁcation (SRC). Kambo and

Yerpude [5] distinguished the variety of Basmati Rice Grain

using K-NN and Principal Component Analysis (PCA), after

preprocessing the images with smoothing and segmentation

techniques. To improve the classiﬁcation accuracy, Rad

et al. [6] presented a rice classiﬁcation algorithm with

optimal morphological features and back propagation neural

network-based (BPNN). Here, 18 morphological features

were extracted, and 6 features were selected. Silva and

Sonnadara [7] combined neural network (NN) with PCA

to classify the rice seed varieties. 34 features were extracted

by some pre-processing operations before PCA was applied to

perform dimensionality reduction, and one individual neural

network was created for each feature set.

With the widespread application of deep learning and their

excellent performance in image tasks, more and more rice

classiﬁcation work starts to adopt deep neural networks. Lin et

al. [8] proposed a model using convolutional neural networks

(CNN) for rice kernel classiﬁcation which reached a 99.52

accuracy. However, they found the accuracy of classiﬁcation

is closely related to the preprocessing effect of the image,

which was done with image enhancement operations, such

as re-scaling, mean subtraction, and feature standardization.

Qiu et al. [9] presented a hyperspectral-CNN based rice

variety classiﬁcation algorithm. The rice image captured from

the hyperspectral system was pre-processed with a wavelet

transform and image segmentation process. 100-3000 rice

samples were used to build KNN, SVM, and CNN models.

They found CNN outperformed SVM and KNN. Similar

to this work, Chatnuntawech et al. [10] provided a rice

classiﬁcation algrithom for identiﬁcation with the synergy

between hyperspectral imaging and deep CNN. They found

that the classiﬁcation effect has been signiﬁcantly improved

with the deepening of the number of neural network layers.

Patel proposed two methods for rice types classiﬁcation. One

used CNN with segmented rice images as input. Another used

a pretrained VGG-16 model and transfer learning to achieve a

better result. Kiratiratanapruk et al. [3] applied deep learning

to detect and identify rice disease in images. They conducted

experiments with 4 models namely Faster R-CNN [11],

RetinaNet [12], YOLOv3 [13] and Mask RCNN [14].

Rice quality analysis:

Different solutions have been

applied on rice grain analysis. These approaches can broadly

be classiﬁed into geometric, statistical, and machine learning.

Geometric approaches consider morphological features as

key factors to analysis. Ajay et al. [2] used shape descriptors

and geometric features to determine quantity of broken

kernels among milled rice samples. Asif et al. [15] used

morphological features to determine the quality of ﬁve

types of rice grains after a grain classiﬁcation. Mahale and

Korde [16] applied image processing techniques to grade

and evaluate rice grains based on grain size and shape, such

as length, width and their ratio. In contrast, Ali et al. [17]

proposed an low cost solution for rice quality analysis based

on more features. They computed the average length, average

width, the area and number of small rice grains, medium rice

grains and broken rice grains to analysis rice quality.

Statistical approaches primarily focus on summarizing data

and making inferences according to the population. Mahajan

and Kaur [18] proposed a method of quality analysis for three

types of Indian Basmati rice grains including normal grains,

long grains and small grains. They applied morphological

closing and opening operations and top-hat transformation

to calculate the length of the major and minor axes of rice.

The rice was graded by analyzing histograms.

With the successful application of machine learning and

its excellent performance, Agustin and Oh [19] proposed an

automatic quality evaluation framework for milled rice kernels,

in which a probabilistic neural network (PNN) classiﬁer is

used to detect defective rice including head rice, broken, and

brewer kernels. At the same time, a linear regression model

was developed for estimating individual kernel weight with

a given blob area. [20] applied neural network and image

processing approach for rice grain identiﬁcation and grading

on three varieties of Indian rice. Rice images, obtained from

a ﬂatbed scanner, were pre-processed with several image

smoothing operations. The length, width, and perimeter of

the rice grains were extracted and input neural network to do

classiﬁcation. It was able to accurately classify rice into sound,

cracked, chalky, broken and damaged kernels. Ngampak and

Piamsa-Nga [21] proposed a method for ﬁnely classifying

broken rice into small broken, broken, big broken and head

rice. They used Least-Square Support Vector Machine (LS-

SVM) with Radius Basis Function (RBF) kernel as their

classiﬁer. Kaur et al. [22] divided rice kernels into four

grades using Multi-Class SVM. They categorized rice into

head rice, broken rice and brewers according to the kernel

shape, length and chalkiness.

While there are many different methods for detecting rice

with different defects, none of them have classiﬁed rice with

dual properties.

III. METHOD

In this section, the rice kernel quality evaluation problem

will be formulated ﬁrst. Then the hardware setup for rice

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Vision-BasedDefectClassicationandWeightEstimationofRiceKernelsXiangWang1,KaiWang1,XiaohongLi1,andShiguoLian1.AbstractRiceisoneofthemainstaplefoodinmanyareasoftheworld.Thequalityestimationofricekernelsarecrucialintermsofbothfoodsafetyandsocio-economicimpact.Thiswasusuallycarriedoutbyqualityinspecto...

展开>> 收起<<

Vision-Based Defect Classiﬁcation and Weight Estimation of Rice Kernels Xiang Wang1 Kai Wang1 Xiaohong Li1 and Shiguo Lian1.pdf

共9页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Vision-Based Defect Classiﬁcation and Weight Estimation of Rice Kernels Xiang Wang1 Kai Wang1 Xiaohong Li1 and Shiguo Lian1

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: