Double Attention-based Lightweight Network for Plant Pest Recognition Sivasubramaniam Janarthan Selvarajah Thuseethan Sutharshan Rajasegarar

2025-05-03 0 0 538.61KB 14 页 10玖币

侵权投诉

Double Attention-based Lightweight Network for

Plant Pest Recognition

Sivasubramaniam Janarthan, Selvarajah Thuseethan, Sutharshan Rajasegarar,

and John Yearwood

Deakin University, Geelong, VIC 3220, Australia

{jsivasubramania,tselvarajah,srajas,john.yearwood}@deakin.edu.au

Abstract. Timely recognition of plant pests from ﬁeld images is sig-

niﬁcant to avoid potential losses of crop yields. Traditional convolu-

tional neural network-based deep learning models demand high compu-

tational capability and require large labelled samples for each pest type

for training. On the other hand, the existing lightweight network-based

approaches suﬀer in correctly classifying the pests because of common

characteristics and high similarity between multiple plant pests. In this

work, a novel double attention-based lightweight deep learning architec-

ture is proposed to automatically recognize diﬀerent plant pests. The

lightweight network facilitates faster and small data training while the

double attention module increases performance by focusing on the most

pertinent information. The proposed approach achieves 96.61%, 99.08%

and 91.60% on three variants of two publicly available datasets with

5869, 545 and 500 samples, respectively. Moreover, the comparison re-

sults reveal that the proposed approach outperforms existing approaches

on both small and large datasets consistently.

Keywords: Plant Pest Recognition ·Double Attention ·Lightweight

Network ·Deep Learning

1 Introduction

Plant pests cause severe damage to crop yields, resulting in heavy losses in food

production and to the agriculture industry. In order to reduce the risk caused

by plant pests, over the years, agricultural scientists and farmers tried various

techniques to diagnose the plant pests at their early stage. Although many so-

phisticated automatic pest recognition algorithms have been proposed in the

past, farmers continue to rely on traditional methods like manual investigation

of pests by human experts. This is mainly because of poor classiﬁcation abil-

ity and limited in-ﬁeld applicability of automatic pest recognition systems [1].

Diﬀerent plant pests share common characteristics, which makes automatic pest

recognition a very challenging task and hence traditional handcrafted feature

extraction based approaches often failed to correctly classify pests [2]. While

conventional deep learning-based techniques achieved benchmark performances

in pest classiﬁcation, they have limited usage with resource constraint devices

arXiv:2210.09956v1 [cs.CV] 4 Oct 2022

2 S. Janarthan et al.

due to their high computational and memory requirements [3]. Large labelled

data requirement for training is another ﬂip side of conventional deep learning

techniques [4]. Recently proposed semi-supervised learning of deep networks is

also not ideal for this problem as they frequently demonstrate low accuracies

and produce unstable iteration results.

In order to prevent the deep model from overﬁtting, it is also paramount

to provide suﬃcient data during the training phase [5]. However, constructing

a large labelled data in the agriculture domain, especially for plant pests, re-

quires not only high standard of expertise but also time-consuming. Moreover,

inaccurate labelling of the training data produces deep models with reduced re-

liability. Few-shot learning concept is proposed simply by replicating humans’

ability to recognize any objects with the help of only a few examples [6]. Few-

shot learning has gained popularity across various domains as it can address

the classiﬁcation task with a few training samples. In few-shot learning, the

classiﬁcation accuracy increases as the number of shots grow. However, a ma-

jor limitation of few-shot learning is that the prediction accuracy drops when

the number of ways increases [7]. Directly applying the classiﬁcation knowledge

learned from meta-train classes to meta-test classes is mostly not feasible, which

is another fundamental problem of this approach [8].

Constructing decent-performing models with the reduced number of train-

able parameters by downsizing the kernel size of convolutions (e.g., from 3 ×3

to 1 ×1 as demonstrated in [9]) is a signiﬁcant step towards the development

of lightweight networks. In recent years, lightweight deep network architectures

have gained growing popularity as an alternative to traditional deep networks

[10,11,12,13]. The MobileNets [14] and EﬃcientNets [15] families are thus far two

most widely used lightweight networks. Several lightweight deep network-based

techniques have also been proposed for real-time pest recognition [35,36]. The

lightweight architectures however suﬀer to reach the expected level of classiﬁca-

tion accuracy as they are essentially developed for faster and lighter deployment

by sacriﬁcing the performance.

Considering this issue, a novel high-performing and lightweight pest recog-

nition approach is proposed in this study, as illustrated in Figure 1. While pre-

serving the lightweight characteristic of the deep network, a double attention

mechanism is infused to enhance the classiﬁcation performance. As the attention

closely imitates the natural cognition of the human brain, the most inﬂuential

regions of the pest images are enhanced to learn better feature representations.

Notably, attention-aware deep networks have shown improved performances in

various classiﬁcation tasks [18,19]. The key contributions of this paper are three-

fold:

–A novel lightweight network-based framework integrated with a double atten-

tion scheme is proposed for enhancing the in-ﬁeld pest recognition, especially

using small training data.

–A set of extensive experiments were conducted under diverse environments

to reveal the feasibility and validate the in-ﬁeld applicability of the pro-

posed framework. To organize diverse environments, three publicly available

datasets consisting of small to large number of pest samples are utilized.

Lightweight Network for Plant Pest Recognition 3

1 2

3 4

5 6

11 12 14 15 . . . . . . . .

. . . .

112x112

56x56

28x28

14x14

7x7

A_1before A_1after

A_2after A_3final

A_2before

Input

Pest Classes

Fig. 1. The overall architecture of the proposed double attention-based lightweight

pest recognition framework. The layers are labelled to match with the layer numbers

indicated in Table 1. Some layers of the proposed architecture are avoided in this

diagram for brevity. The A1bef ore ,A1af ter ,A2bef ore ,A2af ter and A3f inal are the

activation maps generated before and after respective layers as indicated.

–A comparative analysis is performed to show the superiority of this frame-

work over existing state-of-the-art lightweight networks that are often used

for pest recognition methods available in the literature.

The remainder of this paper is constructed as follows. The recent advance-

ments of deep learning based pest recognition approaches are given in Section 2.

Section 4 provides comprehensive details on the proposed framework and Section

5 concludes the paper with some future directions.

2 Related work

Eﬀective pest recognition is essential for preventing the spread of crop diseases

and minimizing economic losses in relation to agriculture. Over the years, re-

searchers invested considerable eﬀort to develop pest recognition techniques.

Probe sampling, visual inspection and insect trap are some of the widely used

manual pest recognition approaches that are still the farmers’ favourites when

it comes to in-ﬁeld tasks [20]. Early studies targetted the sounds emitted by the

pests to perform the classiﬁcation, but the paradigm has quickly shifted to digi-

tal images in the last decade [21]. As this paper reviews a few recently proposed

prominent pest recognition works, readers are recommended to read [22] and

[23] for a comprehensive pest recognition literature.

With the recent development of deep learning, the automatic pest recogni-

tion techniques have become a rapidly growing agricultural research direction

[24,25,26]. Diﬀerent deep convolutional neural networks (CNNs) (i.e., state-of-

the-art and tailor-made networks) and capsule networks are prominently used

to develop high-performing pest recognition methods. In [27], the pre-trained

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：

DoubleAttention-basedLightweightNetworkforPlantPestRecognitionSivasubramaniamJanarthan,SelvarajahThuseethan,SutharshanRajasegarar,andJohnYearwoodDeakinUniversity,Geelong,VIC3220,Australiafjsivasubramania,tselvarajah,srajas,john.yearwoodg@deakin.edu.auAbstract.Timelyrecognitionofplantpestsfromeldima...

展开>> 收起<<

Double Attention-based Lightweight Network for Plant Pest Recognition Sivasubramaniam Janarthan Selvarajah Thuseethan Sutharshan Rajasegarar.pdf

共14页,预览3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！

Double Attention-based Lightweight Network for Plant Pest Recognition Sivasubramaniam Janarthan Selvarajah Thuseethan Sutharshan Rajasegarar

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: