Toward Objective Variety Testing Score Based on Computer Vision and

Unsupervised Machine Learning: Application to Apple Shape

Mouad Zine-El-Abidine

, Helin Dutagaci

, Pejman Rasti

1,3

Maria Jose Aranzana

, Christian Dujak

and David Rousseau

Universit

e d’Angers, LARIS, UMR INRAe, IRHS, France

Eskisehir Osmangazi University, Department of Electrical-Electronics Engineering, Eskisehir, Turkey

CERADE,

Ecole d’Ing

enieur Informatique et Pr

evention des Risques (ESAIP), Angers, France

Centre de Recerca en Agrigen

omica: Cerdanyola del Valles, Catalunya, Spain

Keywords:

Variety Testing, Variety Classiﬁcation, Shape Description, Apple Shape.

Abstract:

While precision agriculture or plant phenotyping are very actively moving toward numerical protocols for ob-

jective and fast automated measurements, plant variety testing is still very largely guided by manual practices

based on visual scoring. Indeed, variety testing is regulated by deﬁnite protocols based on visual observation

of sketches provided in ofﬁcial catalogs. In this article, we investigated the possibility to shortcut the human

visual inspection of these sketches and base the scoring of plant varieties on computer vision similarity of

the ofﬁcial sketches with the plants to be inspected. A generic protocol for such a computer vision based

approach is proposed and illustrated on apple shape classiﬁcation. The proposed unsupervised algorithm is

demonstrated to be of high value by comparison with classical supervised and self supervised machine and

deep learning if some rescaling of the sketches is performed.

1 INTRODUCTION

Plant variety testing refers to the activity of assessing

new varieties of plants before registering them in an

ofﬁcial authorized catalog. The test can assess char-

acteristics such as appearance, resistance to various

stresses and agronomical value. So far variety testing

is mostly based on human visual inspection. For scor-

ing of key characteristics, experts follow guidelines

in the form of either written instructions or visual ref-

erence sketches. Such guidelines are usually deliv-

ered in a catalog by an ofﬁcial variety testing organ-

ism, such as The International Union for the Protec-

tion of New Varieties of Plants (UPOV). Trait descrip-

tors are also scored by breeders and germplasm cura-

tors, who follow in-house or published/recommended

guidelines. Two limitations of manual assessment by

experts are that the rating is non-objective and that the

process is slow and time-consuming. The objective of

this article is to propose a fast and objective generic

protocol adapted to the speciﬁc requirement of variety

testing with the help of computer vision.

Computer vision is now widely used in plant phe-

notyping and agriculture (Mahlein, 2016; Li et al.,

2020; Dhanya et al., 2022) and especially for fruit

phenotyping (Bhargava and Bansal, 2021). Computer

vision has been only recently explored in the domain

of variety testing (Couasnet et al., 2021; El Abidine

et al., 2020; Zine-El-Abidine et al., 2021; Garbouge

et al., 2021b; Garbouge et al., 2021a; Koklu et al.,

2021; Meng et al., 2022; Cao et al., 2022). This ar-

ticle is a new contribution in this recent application

domain of computer vision to variety testing.

One can distinguish two approaches in computer

vision (Szeliski, 2022). The traditional approach is

the model-based approach where the characteristics

of the objects to be analyzed in the scene are trans-

lated into geometrical shapes. Features, i.e. set of

numbers, are then computed on these geometrical

shapes and the computer takes decision in this feature

space. The current leading approach is the data-driven

approach where the features space and the decision

making are built, without the help of a mathematical

expert to translate shape into features, in a supervised

manner based on training data set (part of the data

on which the expect output of the model, i.e. ground

truth, is manually established) to produce a model that

will then be used for inference on unseen data.

Zine-El-Abidine, M., Dutagaci, H., Rasti, P., Aranzana, M., Dujak, C. and Rousseau, D.

Toward Objective Variety Testing Score Based on Computer Vision and Unsupervised Machine Learning: Application to Apple Shape.

DOI: 10.5220/0012549700003720

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 4th International Conference on Image Processing and Vision Engineering (IMPROVE 2024), pages 15-22

ISBN: 978-989-758-693-4; ISSN: 2795-4943

Coupling computer vision with data-driven super-

vised machine learning techniques has been proven

to be successful for many phenotyping tasks (Benos

et al., 2021; Li et al., 2020; Pathan et al., 2020).

However, computer vision models produced via su-

pervised machine learning depend on the training data

set. Consequently, there are some constraints on the

training data set to ensure a good generalization on

the unseen data set. First, the training data set has to

be representative of the unseen data to be tested. Sec-

ond, the size of the training data set has to be large

enough to ﬁt with good generalization properties the

parameters to be adapted in the models. When us-

ing Deep learning, the most powerful and common

tools in data driven approach, the typical minimal size

of the training data set is some few thousands or in-

stances at least. These basics of supervised machine

learning clearly indicate that this approach is not suit-

able to mimic the current protocol for variety testing.

Indeed, in variety testing experts establish their rat-

ings in reference to the visual comparison between a

very small set of sketches (typically some units) and

the inspected real plants. These sketches cannot con-

stitute a large enough training set. Also, since the

experts can disagree on their ratings due to subjec-

tive interpretations of the ofﬁcial sketches, supervised

machine learning using ground truth provided by an

expert may embed some bias to their inference mod-

els and therefore not be representative of the unseen

data to be tested.

Supervised machine learning appears not suitable

for variety testing based. In this work, we investigate

the possibility to directly use reference sketches in an

ofﬁcial catalog for quantitative matching with images

of plants to be assessed (see Fig. 4). We test this ap-

proach, which is novel in the context of variety test-

ing, on the problem of apple shape assessment. The

closest related problem in computer vision is sketch-

based image retrieval (SBIR), where the objective is

the retrieval of related images from a data base given a

sketch query (see (Zhang et al., 2019) for a recent re-

view). The SBIR method combines information from

both datasets (sketches and RGB images) for a high

accuracy image retrieval. Rather than image retrieval,

we target classiﬁcation of RGB images of apples by

quantitatively matching them with catalog sketches.

As another ﬁeld of related work, shape-based clas-

siﬁcation and grading of fruits based on supervised

machine learning has been widely studied in the liter-

ature (Ishikawa et al., 2018; Jana and Parekh, 2017;

Kheiralipour and Pormah, 2017; Hu et al., 2018; Ireri

et al., 2019; Li et al., 2019). By contrast, while we

use classical shape features to characterize the shape

of apples, we do not promote a supervised machine

learning techniques here. To the best of our knowl-

edge, there exists no previous work for development

of a sketch-based classiﬁcation tool in the context of

variety testing.

The article is structured in the following way.

We ﬁrst present the material and methods used for

the apple use case chosen to illustrate our a sketch-

based classiﬁcation tool. For fair comparison we will

naturally compare our approach either in terms of

performance and energy consumption with some of

the state-of-the-art supervised or self-supervised deep

learning methods. We demonstrate and discuss the

domain of superiority of our approach and its generic

interest for other use cases in variety testing conclude.

2 MATERIALS AND METHODS

2.1 Reference Sketches

The apple shape classiﬁcation tool currently follows

the UPOV rules. In this variety testing framework,

experts inspect cut apple shapes by comparing them

to the reference sketches in the ofﬁcial variety test-

ing catalog (see Fig. 1). Three main classes are used

to designate the shape of apples: Flat, Globose and

Oval. For each category, there are sub-categories

such as Flat-Globose, Oblong and Ellipsoid. In this

study, we target classiﬁcation of apple images into

three broad categories: Flat, Globose and Oval.

2.2 Data Set

The image acquisition procedure is shown in Fig. 2.

As image acquisition procedure, apples from the Ref-

pop population (Jung et al., 2020)) are cut along their

medial axis, placed on their ﬂat, freshly cut side in

groups of 6 on an HP Scanjet Pro 4500 fn1 with max-

imum resolution of 1200 x 1200 dpi. Since the con-

trast between the apples and the background is strong,

we used simple thresholding on the brightness chan-

nel of HSB color space to segment the apples from the

background. The bounding box of each individual ap-

ple is obtained through connected component analy-

sis. A simple edge detection via Sobel ﬁlter is applied

to produce a binary image highlighting the boundaries

of the apples (see Fig. (4)).

Overall, 1821 images were acquired and classiﬁed

independently by two experts. Figure 3 shows the

distribution of the classes (Flat, Globose and Oval)

corresponding to the annotations of the two experts.

Three sets of class labels resulted from this annota-

tion. The class labels provided by expert 1, by ex-

pert 2 and a subset labeled in agreement by both ex-

IMPROVE 2024 - 4th International Conference on Image Processing and Vision Engineering

Figure 1: Reproduced and modiﬁed view of the ECPGR catalog. Apple shape sketches in the catalog of variety testing. The

classes considered in this work are highlighted in the red rectangle.

Figure 2: Acquisition device: HP Scanjet Pro 4500 fn1.

perts (600 images) were kept. This quantiﬁes the

inter-variability between expert and the current con-

sequences of subjective rating. This also shows the

intrinsic difﬁculty of the visual tasks raised to experts

in variety testing where experts only agree in 30% of

the cases.

2.3 Shape Descriptors

In this article, since our aim is not to provide new

features to characterize apple shape but rather to in-

vestigate the possibility to use the reference sketches

for apple shape classiﬁcation, we used some exist-

ing shape descriptors (Ishikawa et al., 2018; Ghazal

et al., 2021). We used chain-code histogram, ellip-

tical Fourier descriptors (Kuhl and Giardina, 1982)

and Frechet’s Ratio with the following hyperparam-

eters. The connectivity for the chain-code was take to

8. The 10 ﬁrst harmonics of elliptical Fourier descrip-

tors were kept. All features were normalized to 1 to

allow the use of Euclidean distance in the produced

feature space.

2.4 Experiments

We evaluated two approaches for apple classiﬁca-

tion: i) reference-based classiﬁcation approach and

ii-) three model-based classiﬁcation approaches, one

based on support vector machine (SVM), one based

on supervised deep neural network and one based on

self-supervised deep neural network. The details are

provided in the following subsections.

Toward Objective Variety Testing Score Based on Computer Vision and Unsupervised Machine Learning: Application to Apple Shape

Figure 3: Histogram of the distribution of classes (Flat, Globose and Oval) assigned by experts.

2.4.1 Reference-Based Classiﬁcation Approach

In this experiment, we mimic the way variety testing

experts use catalogs. We perform a multi-class classi-

ﬁcation by computing the Euclidean distance between

features of cut apple query image and features from

a reference instance of each class: one for Flat, one

for Globose and one for Oval. A visual abstract of the

pipeline is given in Fig. (4). Three types of references

were tested i) the original sketches from the ofﬁcial

catalog of ECPGR, ii) the contours from real apples

each representing a class, iii) The ECPGR sketches

rescaled.

First, we considered the original sketches from the

ofﬁcial catalog of ECPGR. As the second option, in-

stead of the sketches provided in a catalog, we consid-

ered the contours of representative apples as reference

shapes. The representative apples are chosen from the

dataset of real apple images. For each class, the apple

with the aspects ratio closest to the class average is se-

lected as the class representative. As the third option,

we modiﬁed ECPGR sketches such that the aspect ra-

tio of each reference sketch becomes equal to the cor-

responding class average. This operation bridges the

gap between the aspect ratios of the sketches and the

distribution of the aspect ratio of the apple variety to

be tested.

The average aspect ratio was computed in the fol-

lowing unsupervised way. We assigned the centroids

of each cluster as the class average aspect ratio to the

corresponding class knowing the ordinal relation be-

tween classes. This sorting was obviously not perfect

(otherwise the task would have been done).

2.4.2 Alternative Classiﬁcation Approaches

In this approach, we do not follow the ECPGR cat-

alog and rather adopt supervised machine learning

techniques. Three models are trained to classify the

shape of images based on a training set composed

of annotated RGB images. Since our goal is not to

claim optimal performances but rather to provide a

comparison with our proposed reference-based ap-

proach, we selected a basic machine learning model

(an SVM with a linear kernel) and two deep learn-

ing algorithms including supervised (Kamilaris and

Prenafeta-Bold

u, 2018; Koirala et al., 2019) and self-

supervised deep learning (G

uldenring and Nalpan-

tidis, 2021). The well-known VGG16 (Simonyan and

Zisserman, 2014) and SimSiam (Chen and He, 2021)

models have been implemented for supervised and

self-supervised deep learning models. To quantify

the sensitivity to the choice of the data reserved for

the training, multiple runs of the classiﬁcation exper-

iment were conducted for various values of the train-

test split of the data set and 10-fold cross-validation.

The average value and standard deviation of the per-

formances of classiﬁcation were recorded. To quan-

tify the inter-variability between annotating experts,

the experiment was repeated with labels provided by

separated labels provided by the two experts and with

the curated data set containing apples with agreed la-

bels only.

2.4.3 Metric

All the classiﬁcation experiments were evaluated us-

ing the accuracy metric

ACC =

T P + TN

T P + T N + FP + FN

. (1)

To rely on this metric, the classes Flat, Globose and

Oval, on both experts’ annotated datasets and cured

dataset were balanced with 200 images for each class.

IMPROVE 2024 - 4th International Conference on Image Processing and Vision Engineering

Figure 4: Illustrative pipeline for the Reference-based classiﬁcation approach: (a1): reference sketches Dataset, (a2): an apple

RGB image to be classiﬁed. (b1): rescaling of the aspect ratio of sketches for each class. (b2): edge detection. (c) rescaled

sketches. (d): shape features extraction. (e): similarity measure. (f): classiﬁcation results; apple RGB image classiﬁed based

on the minila distance.

3 RESULTS

3.1 Reference-Based Classiﬁcation

Approach

Table 1 provides quantitative evaluation of reference-

based approach results. One can observe low accu-

racy when the sketches from the ECPGR catalog are

used as reference. A noticeable improvement of 9

to 13 % of accuracy is brought when these reference

sketches are rescaled. The best performance is ob-

tained when the class representatives are selected as

a reference. However, it is to be noticed that the

gain of performance is only of a 2 to 4 % by com-

parison with the unsupervised rescaling of the ofﬁcial

sketches. The performance culminates at 77% of ac-

curacy when the data are curated. We reproduced the

experiment after withdrawing the intermediate class

Globose. Results given in Table 2 show similar trends

as in Table 1 but with much higher accuracy around

95%. We still observe a gain of 2 to 3% after rescal-

ing the reference sketches of the ECPGR catalog with

our proposed approach.

3.1.1 Comparison with Alternative Approaches

Figures 5 shows the performance of the three model-

based approaches as a function of the ratio of the size

of the test set to the size of the training set. As triv-

ially expected, the performances drop progressively

with increasing standard deviation when the amount

Table 1: Measuring accuracy (ACC) of reference-based

approach on expert 1 (E1), expert 2 (R2) and curated

data (CD), to all types of sketches. Ref stands for refer-

ence sketch, Classrep for class representatives sketches and

Rescaled for rescaled reference sketches.

Sketches % ACC

% ACC

Ref 58% 55% 60%

Classrep 69% 67% 77%

Rescaled 67% 63% 73%

Table 2: Same as in Table 1 but with only two classes Flat

and Oval.

Sketches % ACC

% ACC

Ref 94% 90% 95%

Classrep 97% 96% 98%

Rescaled 97% 95% 97%

of data in the training data set reduces. The plateau

of performance of the SVM-based method for a low

test/training ratio is around 93% for the curated data

sets but drops to around 43% with a huge standard

deviation of 14% when only one instance is kept.

The same behavior could also be seen in the perfor-

mance curved yield by the deep learning and self-

supervised learning approaches. The big gap (around

10% to 20%) in terms of accuracy between the

SVM-based approached and supervised Deep learn-

ing and self-supervised learning approaches shows

that Deep learning and self-supervised learning ap-

proaches are more data-dependent than machine

learning approaches based on handcrafted-features.

Toward Objective Variety Testing Score Based on Computer Vision and Unsupervised Machine Learning: Application to Apple Shape

These results should be compared with the reference-

based classiﬁcation approach explained in the previ-

ous subsection, where only one image per class was

used. It also has to be mentioned that deep learning

and self-supervised learning approaches have a high

computational cost which affects the environment and

global warming. We estimate the amount of carbon

dioxide (CO2) produced by our computing resources

used to execute these two methods (Schmidt et al.,

2021). Our estimation showed that these computa-

tions consumed around 69 kWh, which equaled about

27 kg (59.5 pounds) of CO2 emissions.

The results of predictions on test set of cu-

rated data, using the same model-based approaches

and reference-based approach toward reference

sketches, centers representatives and rescaled refer-

ence sketches, are presented in Fig. 5. It is impor-

tant here to recall that the reference-based approach is

purely unsupervised and therefore fully automatic in

the case of testing of a new variety while the model-

based approach require labor-intensive annotation of

the newly introduced data set.

The performance of the reference-based classiﬁ-

cation is found to be stable with the amount of data

used to compute the rescaling aspect ratio and outper-

forms the SVM-based approach when less than 30%

of the data sets are not annotated. This experiment

was carried out again while withdrawing the inter-

mediate Globose class with similar results shown in

Fig. 6. The difference of plateau of performance be-

tween the SVM-based approach and the reference-

based approach vanishes, and there is here no clear

advantage in annotating the images to train a model.

The current approach based on sketches can directly

be automated with the unsupervised approach pro-

posed in this work.

Figure 5: Prediction curves of test set of curated data via

model-based classiﬁcation, CNN and SSL and reference-

based classiﬁcation using reference sketches, centers repre-

sentatives and rescaled reference sketches, after training on

curated dataset.

Figure 6: Same as in Fig. 5 with only two classes Flat and

Oval.

4 DISCUSSION

As illustrated in Fig. (7), variety testing catalogs may

differ from one country to another, as well as between

germplasm curators and breeders, or the sketches of

reference may evolve. This situation may cause dif-

ﬁculty of comparison of the results over time or even

communication problems between countries not shar-

ing the same references. The instance-based approach

described in this article may actually serve to decipher

this Tour of Babel problem. As shown in Table 3,

we investigated the possibility of automatic transla-

tion of one catalog to another. To this purpose, we

provided the nearest reference sketches to a query ref-

erence sketch from a different catalog.

In Table 3, some catalogs categories are found to

be perfectly matching with each others. In other cases

the designation used in one catalog does not match

with the designation of another catalog. This shape-

based translation based on pure objective features en-

ables to overcome the semantic gap that the multi-

plicity of catalogs may cause. Consequently, our ap-

proach may not only be used to provide an objective

tool conforming the current variety testing practices,

it can also be used when judgements based on dif-

ferent visual references need to be shared. It is here

again important to stress that this translation from one

catalog of reference to another, would not be directly

accessible with supervised machine learning. Indeed

this would require to train models with annotation

provided by experts using different catalogs. Then it

would require to compare this subjective rating result

on testing data. By contrast with the unsupervised

instance-based model proposed in this article, trans-

lation from one catalog to another is almost instanta-

neous since only the objective similarity between ref-

erence sketches need to be computed.

IMPROVE 2024 - 4th International Conference on Image Processing and Vision Engineering

Table 3: Translation between variety testing catalogs. Each line provides K nearest neighbors (K = 6) in the Upov Catalog of

query sketches from ECPRG Catalog. The ﬁrst column gives the query sketches.

ECPGR Catalog Upov Catalog

Flat Flat Oblate Globose-Canonical Ellipsoid Oblong Globose

Globose Globose Globose-Canonical Ellipsoid Oblong Oblate Flat

Oval Globose Ellipsoid Oblong Globose-Canonical Oblate Flat

Ellipsoid Oblong Ellipsoid Globose-Canonical Globose Oblate Flat

Flat-Globose Oblate Flat Ellipsoid Globose-Canonical Oblong Globose

Oblong Globose Oblong Globose-Canonical Oblate Ellipsoid Flat

Figure 7: Two different variety testing catalogs. (a): Cat-

alog delivered by The European Malus GERMPLASM

Workshop (ECPGR 2009). (b): Catalog delivered by UPOV

(2006).

5 CONCLUSION

In this article, we have demonstrated the possibil-

ity of using reference sketches in a variety testing

catalog, to help the transition from pure manual in-

spection toward automated computational practices.

The sketches can serve as references for quantita-

tive matching to classify images of plant instances.

Rescaling of the aspect ratio of the reference sketches

was shown to be helpful to boost the performances of

classiﬁcation. Reference-based approach was shown

to be better suited in variety testing as compared to su-

pervised machine learning approaches since the later

requires intensive manual annotation and therefore

brings no gain of efﬁciency to the current practice of

manual inspection. This work opens several perspec-

tives. Although the proposed methodology was illus-

trated on apple shape evaluation in variety testing, it

could be extended to any of variety testing traits as-

sociated with reference sketches of the ofﬁcial cata-

logs. Some sketches in variety testing correspond to

3D rendered views. In this case the correspondence

with the 2D images would not be as direct as in this

article. It would require to ﬁrst acquire a set of images

in 3D view and then ﬁnd the best match with the pose

of the reference sketch in the catalogue. Concerning

apple, we followed the ofﬁcial protocol and operated

with freshly cut apples. It would be interesting to re-

produce the experiment with uncut apples to ﬁnd if

the variety testing protocol could be adapted uncut

apples for faster and non destructive characterization.

On the side of artiﬁcial intelligence, we demonstrated

the limits of state of the art convolutional neural net-

work either in supervised or self-supervised learning

for limited data set size by comparison with our ap-

proach. It would be interesting to extend the compar-

ison with the recently introduced foundation models

which are expected to perform better with few or even

zero -shot learning.

ACKNOWLEDGEMENTS

We thank Enrique Dapena and Maria Mercedes Fer-

nandez for their contribution in the experienced an-

notation of apple images. CD was partially sup-

ported by the Horizon 2020 Framework Program

of the European Union under grant agreement No

817970 (project INVITE: Innovations in plant vari-

ety testing in Europe to foster the introduction of

new varieties better adapted to varying biotic and abi-

otic conditions and to more sustainable crop man-

agement practices) and by “DON CARLOS AN-

TONIO LOPEZ” Abroad Postgraduate Scholarship

Program, BECAL-Paraguay. We acknowledge sup-

port from the CERCA Programme (“Generalitat de

Catalunya”), and the “Severo Ochoa Programme for

Centres of Excellence” 2016-2019 (SEV-2015-0533)

and 2020-2023 (CEX2019-000902-S) both funded by

MCIN/AEI /10.13039/501100011033.

REFERENCES

Benos, L., Tagarakis, A. C., Dolias, G., Berruto, R., Kateris,

D., and Bochtis, D. (2021). Machine learning in agri-

culture: A comprehensive updated review. Sensors,

21(11):3758.

Bhargava, A. and Bansal, A. (2021). Fruits and vegetables

quality evaluation using computer vision: A review.

Journal of King Saud University-Computer and Infor-

mation Sciences, 33(3):243–257.

Cao, Q., Yang, G., Wang, F., Chen, L., Xu, B., Zhao,

C., Duan, D., Jiang, P., Xu, Z., and Yang, H.

(2022). Discrimination of tea plant variety using in-

situ multispectral imaging system and multi-feature

analysis. Computers and Electronics in Agriculture,

202:107360.

Toward Objective Variety Testing Score Based on Computer Vision and Unsupervised Machine Learning: Application to Apple Shape

Chen, X. and He, K. (2021). Exploring simple siamese rep-

resentation learning. In Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern Recogni-

tion, pages 15750–15758.

Couasnet, G., El Abidine, M. Z., Laurens, F., Dutagaci, H.,

and Rousseau, D. (2021). Machine learning meets

distinctness in variety testing. In Proceedings of the

IEEE/CVF International Conference on Computer Vi-

sion, pages 1303–1311.

Dhanya, V., Subeesh, A., Kushwaha, N., Vishwakarma,

D. K., Nagesh Kumar, T., Ritika, G., and Singh, A.

(2022). Deep learning based computer vision ap-

proaches for smart agricultural applications. Artiﬁcial

Intelligence in Agriculture, 6:211–229.

El Abidine, M. Z., Merdinoglu-Wiedemann, S., Rasti, P.,

Dutagaci, H., and Rousseau, D. (2020). Machine

learning-based classiﬁcation of powdery mildew

severity on melon leaves. In International Confer-

ence on Image and Signal Processing, pages 74–81.

Springer.

Garbouge, H., Rasti, P., and Rousseau, D. (2021a). Deep

learning-based detection of seedling development

from indoor to outdoor. In International Confer-

ence on Systems, Signals and Image Processing, pages

121–131. Springer.

Garbouge, H., Rasti, P., and Rousseau, D. (2021b). Enhanc-

ing the tracking of seedling growth using rgb-depth

fusion and deep learning. Sensors, 21(24):8425.

Ghazal, S., Qureshi, W. S., Khan, U. S., Iqbal, J., Rashid,

N., and Tiwana, M. I. (2021). Analysis of visual

features and classiﬁers for fruit classiﬁcation prob-

lem. Computers and Electronics in Agriculture,

187:106267.

uldenring, R. and Nalpantidis, L. (2021). Self-supervised

contrastive learning on agricultural images. Comput-

ers and Electronics in Agriculture, 191:106510.

Hu, C., Li, D., Song, Y.-Z., Xiang, T., and Hospedales,

T. M. (2018). Sketch-a-classiﬁer: Sketch-based photo

classiﬁer generation. In Proceedings of the IEEE con-

ference on computer vision and pattern recognition,

pages 9136–9144.

Ireri, D., Belal, E., Okinda, C., Makange, N., and Ji, C.

(2019). A computer vision system for defect discrim-

ination and grading in tomatoes using machine learn-

ing and image processing. Artiﬁcial Intelligence in

Agriculture, 2:28–37.

Ishikawa, T., Hayashi, A., Nagamatsu, S., Kyutoku, Y.,

Dan, I., Wada, T., Oku, K., Saeki, Y., Uto, T., Tan-

abata, T., et al. (2018). Classiﬁcation of straw-

berry fruit shape by machine learning. International

Archives of the Photogrammetry, Remote Sensing &

Spatial Information Sciences, 42(2).

Jana, S. and Parekh, R. (2017). Shape-based fruit recog-

nition and classiﬁcation. In International Conference

on Computational Intelligence, Communications, and

Business Analytics, pages 184–196. Springer.

Jung, M., Roth, M., Aranzana, M. J., Auwerkerken, A.,

Bink, M., Denanc

e, C., Dujak, C., Durel, C.-E., i For-

cada, C. F., Cantin, C. M., et al. (2020). The apple

refpop—a reference population for genomics-assisted

breeding in apple. Horticulture research, 7(1):1–16.

Kamilaris, A. and Prenafeta-Bold

u, F. X. (2018). Deep

learning in agriculture: A survey. Computers and elec-

tronics in agriculture, 147:70–90.

Kheiralipour, K. and Pormah, A. (2017). Introducing new

shape features for classiﬁcation of cucumber fruit

based on image processing technique and artiﬁcial

neural networks. Journal of food process engineering,

40(6):e12558.

Koirala, A., Walsh, K. B., Wang, Z., and McCarthy, C.

(2019). Deep learning–method overview and review

of use for fruit detection and yield estimation. Com-

puters and electronics in agriculture, 162:219–234.

Koklu, M., Cinar, I., and Taspinar, Y. S. (2021). Clas-

siﬁcation of rice varieties with deep learning meth-

ods. Computers and electronics in agriculture,

187:106285.

Kuhl, F. P. and Giardina, C. R. (1982). Elliptic fourier fea-

tures of a closed contour. Computer graphics and im-

age processing, 18(3):236–258.

Li, S., Luo, H., Hu, M., Zhang, M., Feng, J., Liu, Y., Dong,

Q., and Liu, B. (2019). Optical non-destructive tech-

niques for small berry fruits: A review. Artiﬁcial In-

telligence in Agriculture, 2:85–98.

Li, Z., Guo, R., Li, M., Chen, Y., and Li, G. (2020). A

review of computer vision technologies for plant phe-

notyping. Computers and Electronics in Agriculture,

176:105672.

Mahlein, A.-K. (2016). Plant disease detection by imag-

ing sensors–parallels and speciﬁc demands for preci-

sion agriculture and plant phenotyping. Plant disease,

100(2):241–251.

Meng, Y., Ma, Z., Ji, Z., Gao, R., and Su, Z. (2022). Fine

hyperspectral classiﬁcation of rice varieties based on

attention module 3d-2dcnn. Computers and Electron-

ics in Agriculture, 203:107474.

Pathan, M., Patel, N., Yagnik, H., and Shah, M. (2020).

Artiﬁcial cognition for applications in smart agricul-

ture: A comprehensive review. Artiﬁcial Intelligence

in Agriculture, 4:81–95.

Schmidt, V., Goyal, K., Joshi, A., Feld, B., Conell, L.,

Laskaris, N., Blank, D., Wilson, J., Friedler, S., and

Luccioni, S. (2021). CodeCarbon: Estimate and Track

Carbon Emissions from Machine Learning Comput-

ing.

Simonyan, K. and Zisserman, A. (2014). Very deep con-

volutional networks for large-scale image recognition.

arXiv preprint arXiv:1409.1556.

Szeliski, R. (2022). Computer vision: algorithms and ap-

plications. Springer Nature.

Zhang, X., Li, X., Liu, Y., and Feng, F. (2019). A survey on

freehand sketch recognition and retrieval. Image and

Vision Computing, 89:67–87.

Zine-El-Abidine, M., Dutagaci, H., Galopin, G., and

Rousseau, D. (2021). Assigning apples to individual

trees in dense orchards using 3d colour point clouds.

Biosystems Engineering, 209:30–52.

IMPROVE 2024 - 4th International Conference on Image Processing and Vision Engineering