Mixing Augmentation and Knowledge-Based Techniques in Unsupervised

Domain Adaptation for Segmentation of Edible Insect States

Paweł Majewski

1 a

, Piotr Lampa

2 b

, Robert Burduk

1 c

and Jacek Reiner

2 d

Faculty of Information and Communication Technology, Wrocław University of Science and Technology, Poland

Faculty of Mechanical Engineering, Wrocław University of Science and Technology, Poland

Keywords:

Augmentation, Domain Adaptation, Instance Segmentation, Edible Insects, Tenebrio Molitor.

Abstract:

Models for detecting edible insect states (live larvae, dead larvae, pupae) are a crucial component of large-scale

edible insect monitoring systems. The problem of changing the nature of the data (domain shift) that occurs

when implementing the system to new conditions results in a reduction in the effectiveness of previously

developed models. Proposing methods for the unsupervised adaptation of models is necessary to reduce the

adaptation time of the entire system to new breeding conditions. The study acquired images from three data

sources characterized by different types of cameras and illumination and checked the inference quality of

the model trained in the source domain on samples from the target domain. A hybrid approach based on

mixing augmentation and knowledge-based techniques was proposed to adapt the model. The ﬁrst stage of the

proposed method based on object augmentation and synthetic image generation enabled an increase in average

from 58.4 to 62.9. The second stage of the proposed method, based on knowledge-based ﬁltering of target

domain objects and synthetic image generation, enabled a further increase in average AP

from 62.9 to 71.8.

The strategy of mixing objects from the source domain and the target domain (AP

=71.8) when generating

synthetic images proved to be much better than the strategy of using only objects from the target domain

(AP

=65.5). The results show the great importance of augmentation and a priori knowledge when adapting

models to a new domain.

1 INTRODUCTION

Edible insects are one of the most promising alterna-

tive sources of novel food. The number of large-scale

edible insect farms is increasing yearly due to the pos-

sibility of obtaining a high-protein product at a rela-

tively low-cost (Dobermann et al., 2017). Edible in-

sect breeding is a good solution for utilizing unused

areas of livestock buildings where animal diseases

such as ASF (African swine fever) previously oc-

curred (Thrastardottir et al., 2021). The need to mea-

sure breeding parameters and detect anomalies, com-

bined with the large-scale nature of breeding, necessi-

tates using a dedicated automated monitoring system.

There have recently been few works regarding

monitoring edible insect breeding related to the Tene-

brio Molitor. (Majewski et al., 2022) proposed a

multi-purpose 3-module system, enabling the detec-

https://orcid.org/0000-0001-5076-9107

https://orcid.org/0000-0001-8009-6628

https://orcid.org/0000-0002-3506-6611

https://orcid.org/0000-0003-1662-9762

tion of edible insect growth stages and anomalies

(dead larvae, pests), semantic segmentation of feed,

chitin, and frass, and larvae phenotyping. The authors

used synthetic images generated from a pool of ob-

jects, signiﬁcantly reducing model development time.

Other works were based on solutions dedicated to sin-

gle issues, e.g. classiﬁcation of larvae segments (Baur

et al., 2022), and classiﬁcation of the gender of pupae

(Sumriddetchkajorn et al., 2015). Undoubtedly, the

results presented in this works demonstrate the fea-

sibility of using methods based on machine learning

and computer vision to inspect edible insect breeding

effectively. However, adapting the developed meth-

ods to new breeding conditions is still an open prob-

lem.

In the literature, we can ﬁnd a signiﬁcant num-

ber of unsupervised model adaptation methods for

the problems of image classiﬁcation (Madadi et al.,

2020), semantic segmentation (Toldo et al., 2020), or

object detection (Oza et al., 2021); however, there

are fewer works in the area of instance segmenta-

tion. Among the most important domain adapta-

380

Majewski, P., Lampa, P., Burduk, R. and Reiner, J.

Mixing Augmentation and Knowledge-Based Techniques in Unsupervised Domain Adaptation for Segmentation of Edible Insect States.

DOI: 10.5220/0011603500003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 5: VISAPP, pages

380-387

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

tion methods are discrepancy-based (Csurka et al.,

2017; Saito et al., 2018), adversarial-based (includ-

ing generative-based) (Yoo et al., 2016; Murez et al.,

2018), reconstruction-based (including graph-based)

(Cai et al., 2019) and self-supervision-based (Khod-

abandeh et al., 2019; Shin et al., 2020). A rela-

tively simple and intuitive approach to domain adap-

tation is pseudo-label-based self-training, which in-

volves training the model for the target domain based

on samples with pseudo-labels representing a predic-

tion of the model trained on labelled samples from

the source domain. An important element in this ap-

proach is prediction ﬁltering.

The pseudo-label-based self-training approach

seems suitable for instance segmentation and even

easier to apply than in object detection. Namely, hav-

ing masks for objects, it is possible to extract them

from images, add them to appropriate object pools

and use them further to generate synthetic images. It

is also easier to perform ﬁltering at the object level, as

it is possible to calculate features for a speciﬁc object.

This work proposed a two-stage hybrid method for

domain adaptation based on using pseudo-labels for

self-training. In 1st stage, it was proposed to expand

the training set of samples through augmentation at

the image and object levels to reduce the overﬁtting

of the model on the source domain. In 2nd stage, ﬁl-

tering of the obtained predictions was carried out us-

ing domain knowledge. An essential contribution of

this work is the study of the importance of creating

the training set in the 1st and 2nd stages, especially

the concept of mixing real and synthetic samples and

mixing samples from the source domain (with real

labels) and the target domain (with pseudo-labels).

In addition, the consequences of using only synthetic

data (no real labelled samples in the training set) on

the model’s performance in cases of inference in and

out of the domain were also examined.

2 MATERIAL AND METHODS

2.1 Problem Deﬁnition

The problem addressed is detection and segmentation

from images of three states of edible insects, namely

(1) live larvae, (2) dead larvae, and (3) pupae. The

samples are in the form of 512x512 images and come

from three sources associated with different types of

recording cameras and lighting, namely (1) CA, (2)

LU, and (3) JA. Examples of samples from the consid-

ered sources, along with the type of objects detected,

are shown in Figure 1.

Figure 1: Examples of samples from the considered

sources: (a) RGB images, (b) types of detected objects.

The main objective of the research was to pro-

pose a suitable domain adaptation method to train the

model on one data source (source domain) with la-

belled samples and make inference on another (target

domain) with unlabelled samples with relatively low

error. The proposed method is expected to reduce the

destructive effect of domain shift on the accuracy of

target domain prediction.

2.2 Data Sources

The samples were acquired using three image acqui-

sition systems, differing in the cameras and lighting

used. The ﬁrst one (CA) was an experimental sta-

tion with a EOS 50D camera (Canon, Tokyo, Japan)

with a resolution of 5184 x 3456 pixels and a zoom

lens. Diffuse white ﬂuorescence lighting was used.

The second (LU) was a data acquisition station pur-

posely built for imaging insects in breeding boxes. It

used a Phoenix PHX120S-CC (LUCID Vision Labs,

Richmond, Canada) camera with a resolution of 4096

x 3000 pixels and a 12 mm focal length lens. Samples

were illuminated with neutral white LEDs in a diffu-

sion tunnel. The third (JA) was a machine vision sys-

tem prepared for industrial implementation for Tene-

brio Molitor breeding. A GOX-12401C-PGE (JAI,

Copenhagen, Denmark) camera was used, with a res-

olution of 4096 x 3000 pixels and a 12 mm lens. In

this case, due to size limitations, LED strips providing

cold white light were used for direct illumination.

2.3 Dataset

A dataset was prepared for the study, containing sam-

ples from various deﬁned sources along with marked

object masks from the deﬁned classes. A total of 15

samples from CA, 29 samples from LU and 36 sam-

Mixing Augmentation and Knowledge-Based Techniques in Unsupervised Domain Adaptation for Segmentation of Edible Insect States

381

ples from JA were labelled. A summary of the la-

belled number of objects can be found in Table 1.

Table 1: The number of objects from deﬁned classes in the

considered image sources.

source type object type no. of objects

live larvae 656

dead larvae 250

pupae 124

live larvae 163

dead larvae 55

pupae 83

live larvae 1247

dead larvae 148

pupae 187

2.4 Data Exploration

For initial data exploration and qualitative evaluation

of domain shift, PCA and visualization of selected

components were performed. The FID (Frechet In-

ception Distance) metric (Heusel et al., 2017) was

also calculated as a measure of the similarity of fea-

tures extracted from images belonging to different

sources. Lower values of the FID metric mean higher

similarity of sample distributions. FID and PCA were

based on a feature vector of length 2048 extracted

from the last pooling layer of the deep convolutional

neural network Inceptionv3 (Szegedy et al., 2015).

Masked images of objects (without surrounding back-

ground) were used for feature extraction.

2.5 Domain Adaptation with Mixing

Augmentation and

Knowledge-Based Techniques

The proposed adaptation method consists of two

stages described in detail in the following sections.

The ﬁrst stage is based on the augmentation of source

domain objects and the generation of synthetic im-

ages. The second stage considers ﬁltering target ob-

jects based on domain knowledge and re-generating

synthetic images using new target domain objects.

The idea scheme of the proposed solution is shown

in Figure 2.

The method for generating synthetic images in-

volved randomly placing objects on the background

image and allowing the simulation of object overlap

in dense scenes. Each generated synthetic image was

associated with an automatically generated label. The

method of generating synthetic images is described

in more detail in (Majewski et al., 2022; Toda et al.,

2020).

Figure 2: Idea scheme of the proposed solution detailing

two stages.

2.5.1 First Stage of Approach with Objects

Augmentation

The basis for training models is a set of real labelled

samples from the source domain. Evaluation results

for a model trained only on a set of real samples (the

only_real method) were used as a reference for the

following proposed approaches.

Object-level augmentation and synthetic image

generation were proposed to extend the source do-

main samples distribution. First, individual objects

were extracted from real images. Then, these objects

were augmented, modifying colour, contrast, sharp-

ness and brightness. The generated objects were then

placed on the background image, obtaining automati-

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

382

cally labelled synthetic images. Examples of the aug-

mented objects and the generated synthetic images are

shown in Figure 2.

Three possibilities for constructing the training

set for 1st stage were identiﬁed. The only_real

method assumes training only on real data, the

real_synthetic method - on real and synthetic data,

and the only_synthetic method - only on synthetic

data. For each setting, the Mask-RCNN (He et al.,

2017) with backbone ResNet-50 (He et al., 2016)

model was trained with default training parameters.

An implementation of the Mask R-CNN model from

the detectron2 (Wu et al., 2019) library was used in

the study. The part related to the 1st stage in Figure 2

shows the real_synthetic approach for creating the

training set.

2.5.2 Second Stage of Approach with

Knowledge-Based Filtering

The ﬁrst component of 2nd stage of the proposed so-

lution is an inference using the model trained in 1st

stage on unlabelled target domain samples. The re-

sulting predictions were treated as pseudo-labels that

needed to be ﬁltered to remove false positive predic-

tions. For ﬁltering, it was used a priori domain invari-

ant knowledge, namely: (1) live larvae are the major-

ity class (see in Table 1), (2) objects from the classes

live larvae and dead larvae are the longest (have the

largest dimension of the longer side of the bound-

ing box), (3) objects from the dead larvae class have

the lowest pixel intensity, (4) objects from the pupae

class have the highest pixel intensity. The proposed

knowledge-based ﬁltering assumes successively:

1. selection of objects with a minimum length of the

longer side of the bounding box d

min

, with a pre-

dicted class live larvae,

2. removal of outliers including mean intensity, size,

and length of the longer side of the bounding box

among the objects extracted in 1st step, obtaining

a distribution of samples representing live larvae,

3. calculation of intensity limits x

min

, x

max

for sam-

ples representing live larvae,

4. selection of objects with intensity values greater

than x

max

, with predicted class pupae,

5. selection of objects with intensity values less than

min

, with predicted class larvae dead.

6. removal of outliers among the objects extracted in

the 4th and 5th steps, obtaining a distribution of

samples representing pupae and dead larvae.

The obtained new samples in the form of target

domain objects and new generated objects after aug-

mentation are then used to generate synthetic data.

In 2nd stage, we have available the following sam-

ple distributions: (1) real from the source domain,

(2) synthetic from the source domain, (3) synthetic

from the target domain. The study investigated the

following training strategies: the "only target domain

samples" strategy assumes training the model only

on synthetic data from the target domain, and the

"mixed source/target domain samples" strategy as-

sumes mixing samples from the source domain and

target domain in the training set. Considering the

"mixed source/target domain samples" strategy, in

all the approaches identiﬁed in 1st stage (only_real,

real_synthetic, only_synthetic), the training set de-

ﬁned in 1st stage is extended with synthetic sam-

ples from the target domain. Figure 2 shows the

"mixed source/target domain samples" strategy with

the real_synthetic variant.

2.6 Evaluation

The proposed methods were evaluated using the av-

erage precision AP

metric, a standard metric for the

evaluation in object detection tasks. The value of the

metric represents the area under the precision-

recall curve after appropriately interpolating the chart

fragments. The AP

metric assumes a threshold value

of intersection over union (IoU) 50% between the true

and predicted bounding box to consider the prediction

signiﬁcant. Details regarding the calculation of the

metric can be found in (Majewski et al., 2022;

Padilla et al., 2020).

For the study, 6 possible cases of out-domain

crossing (source domain → target domain) were

selected, namely CA → LU, CA → JA, LU → CA,

LU → JA, JA → CA, JA → LU. Evaluation for the

out-domain inference cases was carried out for all

samples from the target domain. The AP

values for

in-domain inference were also determined as a refer-

ence. For the in-domain case, the entire dataset was

divided into training data (80% of samples) and test

data. Evaluation was performed on the test set.

3 RESULTS AND DISCUSSION

As part of the data exploration, PCA was performed,

and FID metrics were calculated between sample dis-

tributions. A visualization of the selected components

for samples from all data sources and deﬁned classes

can be found in Figure 3. The calculated FID values

can be found in Table 2.

Based on the results from Figure 3 and Table 2,

it can be seen that objects from the live larvae class

are most similar to each other between distributions.

Mixing Augmentation and Knowledge-Based Techniques in Unsupervised Domain Adaptation for Segmentation of Edible Insect States

383

Figure 3: Selected principal components for domain shift

exploration based on deep features (Inceptionv3).

Table 2: Comparison of calculated FID metrics between

sources based on real samples.

sources object type FID

CA and LU

live larvae 124

dead larvae 166

pupae 144

all (average) 145

CA and JA

live larvae 69

dead larvae 110

pupae 113

all (average) 97

LU and JA

live larvae 97

dead larvae 120

pupae 115

all (average) 111

The FID distances between (CA and JA) and (LU and

JA) distributions are smaller than the distance be-

tween (CA and LU) distributions, which is also con-

ﬁrmed by Figure 3. For the selected components

(PC3 and PC4), samples from JA (blue markers) are

between samples from CA (red) and LU (green).

A comparison of different domain adaptation

methods can be found in Table 3 for the "mixed

source/target domain samples" strategy and in Table 4

for the "only target domain samples" strategy. As

reference values for assessing the quality of domain

adaptation are the results obtained by the models

trained and tested in-domain presented in Table 5.

The results obtained for 1st stage of model adap-

tation (Table 3) prove that the real_synthetic method

(average AP

= 62.9), which assumes the use of both

real and synthetic samples for training, is the most

suitable for use in the problem under consideration.

The use of only synthetic samples (only_synthetic

method, average AP

= 54.4) or only real samples

(only_real method, average AP

= 58.4) may be bet-

ter in special cases (only_synthetic for LU → CA,

only_real for LU → JA, CA → JA), but in gen-

eral (averaged), these approaches achieve smaller

AP values than the real_synthetic method. For

the special cases mentioned above, the differ-

ence between the best-obtained result and the AP

value for the real_synthetic method did not exceed

∆AP

= 4.0. On the other hand, for the LU → CA,

the difference between the AP values for only_real

and real_synthetic was ∆AP

= 14.0, and for the

JA → CA was ∆AP

= 9.4, which is a signiﬁcant dif-

ference in the effectiveness of the models.

Using only synthetic data for model training can

signiﬁcantly speed up the process of developing mod-

els (Majewski et al., 2022); however, based on the

results obtained in this research, we can observe that

this is associated with the risk of losing model accu-

racy. This observation is conﬁrmed by the results af-

ter the 1st and 2nd stages of domain adaptation for

inference out-domain in Table 3 (the only_synthetic

approach was characterized by ∆AP

= 8.5 lower

than the real_synthetic approach in the 1st stage

and by ∆AP

= 4.4 lower AP

in the 2nd second).

The lack of real data in the training set mostly af-

fects the results for in-domain inference in Table 5

(∆AP

= 11.3 difference between only_synthetic and

real_synthetic approaches).

When considering the results separately for each

of the deﬁned classes, it should be noted that objects

of the live larvae class are the easiest to detect (av-

erage AP

after 2nd stage – 81.8) after a domain

change, while objects of the pupae class – the most

difﬁcult (average AP

after 2nd stage – 66.6). This

is consistent with initial conclusions from data explo-

ration based on FID values in Table 2.

Quantitative indicators conﬁrm the importance of

augmentation in 1st stage for the real_synthetic ap-

proach. Additional samples complement the relevant

places in the feature space and can expand the distri-

bution of features for a given class.

Analyzing the results from 2nd stage for the

two proposed strategies in Table 3 for "mixed

source/target samples" strategy and in Table 4 for

"only target samples" strategy, we can conclude that

the "mixed source/target samples" strategy is the most

suitable for creating a training set, which is conﬁrmed

by obtaining an increased AP

from 65.5 to 71.8

compared to the "only target samples" strategy.

A summary of the most important results ob-

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

384

Table 3: Comparison of proposed domain adaptation methods for mixed source/target domain samples strategy.

case type method

stage 1. stage 2. (mixing strategy)

live l. dead l. pup. avg. live l. dead l. pup. avg.

CA → LU

only_real 64.9 58.9 74.7 66.2 79.9 59.7 75.4 71.7

real_synthetic 70.7 61.0 78.0 69.9 82.8 61.7 78.1 74.2

only_synthetic 63.3 54.0 65.5 60.9 82.1 62.0 77.4 73.8

CA → JA

only_real 69.7 50.4 29.2 49.8 75.3 55.1 36.7 55.7

real_synthetic 72.2 38.5 27.6 46.1 76.0 55.0 36.5 55.8

only_synthetic 59.3 28.6 18.6 35.5 77.3 40.8 31.7 49.9

LU → CA

only_real 41.3 58.5 39.5 46.4 79.7 78.1 69.4 75.7

real_synthetic 65.1 68.8 47.2 60.4 80.2 76.4 69.6 75.4

only_synthetic 64.3 69.3 49.9 61.2 79.8 76.7 70.6 75.7

LU → JA

only_real 73.8 53.8 28.7 52.1 83.3 66.9 56.2 68.8

real_synthetic 74.1 45.9 26.2 48.7 84.9 62.3 62.9 70.0

only_synthetic 59.3 31.6 14.7 35.2 83.2 50.5 55.2 63.0

JA → CA

only_real 76.8 67.1 47.8 63.9 84.6 76.4 66.8 75.9

real_synthetic 78.8 73.3 67.7 73.3 82.9 76.3 69.5 76.2

only_synthetic 71.4 73.6 61.1 68.7 78.5 72.8 59.8 70.4

JA → LU

only_real 75.2 68.3 71.5 71.7 83.2 74.3 84.2 80.6

real_synthetic 82.9 71.5 82.7 79.0 84.2 70.6 82.8 79.2

only_synthetic 74.2 60.3 60.3 64.9 79.6 61.9 73.3 71.6

all (summary)

only_real 67.0 59.5 48.6 58.4 81.0 68.4 64.8 71.4

real_synthetic 74.0 59.8 54.9 62.9 81.8 67.1 66.6 71.8

only_synthetic 65.3 52.9 45.0 54.4 80.1 60.8 61.3 67.4

Table 4: Results for the only target samples strategy for the 2nd stage of domain adaptation.

case type method

stage 2. (only target samples strategy)

live larvae dead larvae} pupae average

all (summary)

only_real 76.6 60.9 55.8 64.5

real_synthetic 78.5 59.1 59.0 65.5

only_synthetic 77.5 54.7 54.2 62.2

Table 5: Reference values for domain adaptation as AP

values for in-domain inference.

source type method

live larvae dead larvae pupae average

all (summary)

only_real 86.6 78.8 84.4 83.3

real_synthetic 88.4 81.8 85.2 85.2

only_synthetic 80.0 71.6 70.3 73.9

tained in the study is presented on the radar plot

in Figure 4. In Figure 4 it can be seen that for the

crossings JA → CA, JA → LU, CA → LU, already

1st stage of the proposed method based on augmen-

tation achieves reasonable AP results when using the

real_synthetic approach. The 2nd stage caused a sig-

niﬁcant increase in AP for the crossings LU → JA and

LU → CA. After the two stages of the proposed solu-

tion, the ﬁnal value of the obtained AP values strongly

depended on the target domain. For crossings where

the target domain was JA, the ﬁnal AP values were

the lowest (AP

=55.8, AP

=70). In summary, the

best variation of the proposed method made it pos-

sible to increase the average AP

from 58.4 to 62.9

after 1st stage and to 71.8 after the 2nd stage. To ob-

tain as high AP

values as in-domain trained mod-

els (AP

= 85.2), additional labelling should be per-

formed, especially of objects undetected by models

after the 2nd stage of adaptation. The obtained AP

level between 70 and 80.6 for 5 out of 6 ( except for

CA → JA) types of crossings between domains makes

it possible to improve additional labelling by label

proposals that are predictions of the model obtained

after the 2nd stage.

Mixing Augmentation and Knowledge-Based Techniques in Unsupervised Domain Adaptation for Segmentation of Edible Insect States

385

Figure 4: Comparison of proposed domain adaptation meth-

ods for different cases.

To conﬁrm the good quality of predictions after

domain adaptation, Figure 5 compares the predictions

by the in-domain trained model with the predictions

of the model after domain adaptation for three se-

lected samples from different target domains. For

clarity, Figure 5 shows the predictions only for the

dead larvae and pupae classes.

Figure 5: Comparison of predictions with ground truth for

in-domain and out-domain inference cases.

4 CONCLUSIONS

The proposed two-stage method for domain adapta-

tion made it possible to signiﬁcantly increase the efﬁ-

ciency of object detection (AP

increased from 58.4

to 71.8) when changing the domain without additional

user supervision. The best results were obtained when

the ﬁnal training set consisted of real samples from

the source domain, synthetic samples from the source

domain and synthetic samples from the target domain

(associated with ﬁltered objects from the target do-

main). It conﬁrms the validity of mixing real and syn-

thetic samples in the training set and mixing objects

from the source and target domains. It can also be

concluded from the results that using only synthetic

data when training models can signiﬁcantly reduce

the efﬁciency of the models for both in-domain and

out-domain inference. The study showed the impor-

tance of augmentation techniques and consideration

of a priori knowledge for domain adaptation.

The proposed method is ﬂexible and can be ex-

tended to other classes of objects representing states

of edible insects, e.g., beetles. The method’s exten-

sion would include adding new rules when ﬁltering

the prediction based on a priori knowledge. The de-

veloped solutions will undoubtedly help rapidly adapt

monitoring systems for breeding the Tenebrio Molitor

to new breeding conditions.

Future research should focus on increasing the

quality of synthetic data. An interesting research di-

rection is to develop synthetic images based only on a

priori knowledge independently of a speciﬁc domain.

This approach could obtain a basic model not overﬁt-

ted on a particular domain.

ACKNOWLEDGEMENTS

We wish to thank Mariusz Mrzygłód for developing

applications for the designed data acquisition work-

station. We wish to thank Paweł Górzy

nski and

Dawid Biedrzycki from Tenebria (Lubawa, Poland)

for providing a data source of boxes with Tenebrio

Molitor. The work presented in this publication was

carried out within the project “Automatic mealworm

breeding system with the development of feeding

technology” under Sub-measure 1.1.1 of the Smart

Growth Operational Program 2014-2020 co-ﬁnanced

from the European Regional Development Fund on

the basis of a co-ﬁnancing agreement concluded with

the National Center for Research and Development

(NCBiR, Poland); grant POIR.01.01.01-00-0903/20.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

386

REFERENCES

Baur, A., Koch, D., Gatternig, B., and Delgado, A. (2022).

Noninvasive monitoring system for tenebrio molitor

larvae based on image processing with a watershed al-

gorithm and a neural net approach. Journal of Insects

as Food and Feed, pages 1–8.

Cai, Q., Pan, Y., Ngo, C.-W., Tian, X., Duan, L., and Yao,

T. (2019). Exploring object relation in mean teacher

for cross-domain detection. In Proceedings of the

IEEE/CVF Conference on Computer Vision and Pat-

tern Recognition, pages 11457–11466.

Csurka, G., Baradel, F., Chidlovskii, B., and Clinchant, S.

(2017). Discrepancy-based networks for unsupervised

domain adaptation: a comparative study. In Proceed-

ings of the IEEE International Conference on Com-

puter Vision Workshops, pages 2630–2636.

Dobermann, D., Swift, J., and Field, L. (2017). Opportu-

nities and hurdles of edible insects for food and feed.

Nutrition Bulletin, 42(4):293–308.

He, K., Gkioxari, G., Dollár, P., and Girshick, R. (2017).

Mask r-cnn. In Proceedings of the IEEE international

conference on computer vision, pages 2961–2969.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-

ual learning for image recognition. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 770–778.

Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and

Hochreiter, S. (2017). Gans trained by a two time-

scale update rule converge to a local nash equilibrium.

Advances in neural information processing systems,

30.

Khodabandeh, M., Vahdat, A., Ranjbar, M., and Macready,

W. G. (2019). A robust learning approach to do-

main adaptive object detection. In Proceedings of the

IEEE/CVF International Conference on Computer Vi-

sion, pages 480–490.

Madadi, Y., Seydi, V., Nasrollahi, K., Hosseini, R., and

Moeslund, T. B. (2020). Deep visual unsupervised

domain adaptation for classiﬁcation tasks: a survey.

IET Image Processing, 14(14):3283–3299.

Majewski, P., Zapotoczny, P., Lampa, P., Burduk, R., and

Reiner, J. (2022). Multipurpose monitoring system

for edible insect breeding based on machine learning.

Scientiﬁc Reports, 12(1):1–15.

Murez, Z., Kolouri, S., Kriegman, D., Ramamoorthi, R.,

and Kim, K. (2018). Image to image translation for

domain adaptation. In Proceedings of the IEEE Con-

ference on Computer Vision and Pattern Recognition,

pages 4500–4509.

Oza, P., Sindagi, V. A., VS, V., and Patel, V. M. (2021).

Unsupervised domain adaptation of object detectors:

A survey. arXiv preprint arXiv:2105.13502.

Padilla, R., Netto, S. L., and Da Silva, E. A. (2020). A sur-

vey on performance metrics for object-detection algo-

rithms. In 2020 international conference on systems,

signals and image processing (IWSSIP), pages 237–

242. IEEE.

Saito, K., Watanabe, K., Ushiku, Y., and Harada, T. (2018).

Maximum classiﬁer discrepancy for unsupervised do-

main adaptation. In Proceedings of the IEEE con-

ference on computer vision and pattern recognition,

pages 3723–3732.

Shin, I., Woo, S., Pan, F., and Kweon, I. S. (2020).

Two-phase pseudo label densiﬁcation for self-training

based domain adaptation. In European conference on

computer vision, pages 532–548. Springer.

Sumriddetchkajorn, S., Kamtongdee, C., and Chanhorm, S.

(2015). Fault-tolerant optical-penetration-based silk-

worm gender identiﬁcation. Computers and Electron-

ics in Agriculture, 119:201–208.

Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.,

Anguelov, D., Erhan, D., Vanhoucke, V., and Rabi-

novich, A. (2015). Going deeper with convolutions.

In Proceedings of the IEEE conference on computer

vision and pattern recognition, pages 1–9.

Thrastardottir, R., Olafsdottir, H. T., and Thorarinsdottir,

R. I. (2021). Yellow mealworm and black soldier ﬂy

larvae for feed and food production in europe, with

emphasis on iceland. Foods, 10(11):2744.

Toda, Y., Okura, F., Ito, J., Okada, S., Kinoshita, T., Tsuji,

H., and Saisho, D. (2020). Training instance segmen-

tation neural network with synthetic datasets for crop

seed phenotyping. Communications biology, 3(1):1–

12.

Toldo, M., Maracani, A., Michieli, U., and Zanuttigh, P.

(2020). Unsupervised domain adaptation in semantic

segmentation: a review. Technologies, 8(2):35.

Wu, Y., Kirillov, A., Massa, F., Lo, W.-Y., and Gir-

shick, R. (2019). Detectron2. https://github.com/

facebookresearch/detectron2.

Yoo, D., Kim, N., Park, S., Paek, A. S., and Kweon, I. S.

(2016). Pixel-level domain transfer. In European con-

ference on computer vision, pages 517–532. Springer.

Mixing Augmentation and Knowledge-Based Techniques in Unsupervised Domain Adaptation for Segmentation of Edible Insect States

387