=Paper=
{{Paper
|id=Vol-2696/paper_84
|storemode=property
|title=Enhanced Localization and Classification of Coral Reef Structures and Compositions
|pdfUrl=https://ceur-ws.org/Vol-2696/paper_84.pdf
|volume=Vol-2696
|authors=Kirill Bogomasov,Philipp Grawe,Stefan Conrad
|dblpUrl=https://dblp.org/rec/conf/clef/BogomasovG020
}}
==Enhanced Localization and Classification of Coral Reef Structures and Compositions==
https://ceur-ws.org/Vol-2696/paper_84.pdf
Enhanced Localization and Classification of
Coral Reef Structures and Compositions
Kirill Bogomasov, Philipp Grawe, and Stefan Conrad
Heinrich Heine University, Universitätsstraße 1, 40225 Düsseldorf, Germany
http://dbs.cs.uni-duesseldorf.de
{bogomasov,grawe,stefan.conrad}@hhu.de
Abstract. The automatic annotation of coral images is important for
researching the underwater ecosystem, which is the focus of the Image-
CLEFcoral task. We participated by refining our approaches from the
last years challenge for localization and classification of corals within
images of sea floor. Underwater images bear multiple difficulties which
we tackle with applying image enhancement algorithms. To locate and
classify the corals we applied multiple deep learning approaches and also
revisioned our two-staged algorithm. The results show that deep learn-
ing approaches are the most convincing. Still, the localization of corals is
the most challenging part for us, but we managed to increase our models
performance significantly.
Keywords: Image Segmentation · Image Classification · Object Local-
ization
1 Introduction
Monitoring coral reefs and their health is an important component to understand
effects of the climate change on maritime life [8]. Experts annotate underwater
images, who not only have to deal with the complex morphology of the corals
but also the large number of pictures. Computer vision based localization and
classification of corals seems to be a reasonable solution. Unlike typical datasets
for object detection tasks, underwater images hold more problems regarding the
image quality and thus need very specific features and preprocessing.
In this paper we present the improvements of our approaches which are based
on the last years ImageCLEFcoral [3][5] submission. Additionally we used an-
other deep learning approach, namely RetinaNet [13], since this seemed to be
the most promising. The classical machine learning is revisioned, but still not
compatible with deep learning approaches regarding its performance. Lastly we
implemented a popular suggestion to increase the image quality by preprocessing
the images with algorithms made for underwater photography [7][1].
Overall we increased the performance of our approaches and can provide
more insights, which we present in the following sections.
Copyright c 2020 for this paper by its authors. Use permitted under Creative Com-
mons License Attribution 4.0 International (CC BY 4.0). CLEF 2020, 22-25 Septem-
ber 2020, Thessaloniki, Greece.
2 Related Work
The research of last year coral task can be divided into classical feature engineer-
ing and deep learning approaches. Caridade and Marçal [4] used random forest
classification, based on a selected feature set, consisting of color and texture
features to localize and classify the substrate types. Jaisakthi et al. [10] used a
faster R-CNN to solve this task. Another solution proposal presented by Steffens
at el. [22] is based on a DCNN architecture. Our approach differs from the men-
tioned research. Considering the different properties, distributions and sizes of
the corals, we rely on a combination of both categories of image processing. The
good results substantiate our approach and make it one of the most promising
so far.
3 Data
For the purpose of the task [9][6], a training dataset with 440 images and 12077
annotated substrates, which are labeled with one of 13 substarte types, is pro-
vided. An additional dataset contains 400 raw images, that is used for testing
the predictions while no further information about the images is given to the
participants.
Table 1: Substrate types with their relative frequency in the training set.
Class label Relative frequency
c algae macro or leaves 0.00761463
c fire coral millepora 0.00157259
c hard coral boulder 0.13590465
c hard coral branching 0.09774872
c hard coral encrusting 0.07829829
c hard coral foliose 0.01464989
c hard coral mushroom 0.01845721
c hard coral submassive 0.01638802
c hard coral table 0.00173812
c soft coral 0.46871379
c soft coral gorgonian 0.0074491
c sponge 0.13996027
c sponge barrel 0.01150472
The substrate representatives have a highly imbalanced distribution as shown
in table 1. A random split of the data can lead to an even more disadvantageous
class distribution, since it can exclude the representatives of the rare classes in
one of the sets or increase the impact of frequent classes on the set. We give an
example with a split into train and validation subsets with a ratio of 80 : 20 which
distribution can be seen in table 2. To prevent the tendency of high imbalance,
we propose the following procedure similar to [22].
May P (D) be the relative distribution of classes of the complete image
dataset D. Since both splits of the dataset should have the same distribution of
classes, P (D) is our target distribution. May A and B, with A ∩ B = ∅, be the
two sets after splitting D with the relative distributions P (A) and P (B). The
key idea is to swap images between the two initial random splits A and B to
make P (A), P (B) and P (D) as similar as possible. Let a, b with a ∈ A and b ∈ B
be the two images to swap between A and B. We define A∗ = (A \ {a}) ∪ {b}
and respectively B ∗ = (B \ {b}) ∪ {a}. If the similarity between P (A∗ ) and P (D)
is smaller than the similarity between P (A) and P (D), we swap the items, so
that A = A∗ and B = B ∗ . Since the similarity between P (B ∗ ) and P (D) de-
creases when the similarity between P (A∗ ) and P (D) decreases, this approach
works w.l.o.g. The loop over the images is running until there are no more swaps,
i.e. no swap increases the similarity. To measure this similarity we use Jensen-
Shannon divergence [11]. Since this procedure only converges to a local, but
not global optimum, not every random split ends up in the same balanced split
but consequently the optimal split is not found every time. Further research is
needed to evaluate this approach, in regards of optimizations and metrics. The
results of the balancing algorithm are shown in the table 3.
The Jensen-Shannon divergence between the train and validation set before
swapping is 0.040313, whereas after swapping its divergence is 0.0061. This is
an improvement by a factor of almost 8.
Table 2: Substrate distribution before balancing.
Class label Relative frequency train Relative frequency valid
c algae macro or leaves 0.00799747 0.005827505
c fire coral millepora 0.00126276 0.00271950272
c hard coral boulder 0.12722298 0.167443667
c hard coral branching 0.0967063 0.101787102
c hard coral encrusting 0.07965906 0.0730380730
c hard coral foliose 0.01515311 0.012810513
c hard coral mushroom 0.01894139 0.0167055167
c hard coral submassive 0.01746817 0.01243201
c hard coral table 0.0021046 0.000388500389
c soft coral 0.47606019 0.442113442
c soft coral gorgonian 0.00683995 0.00971250971
c sponge 0.13869304 0.144910645
c sponge barrel 0.01189098 0.101010101
4 Approaches
The ”Coral reef image annotation and localisation task” can be divided into
two tasks. Segmentation of various coral objects from images of sea ground and
classification of those with their specific type of one of the 13 known types of
Table 3: Substrate distribution after balancing.
Class label Relative frequency train Relative frequency valid
c algae macro or leaves 0.00758972 0.00735809
c fire coral millepora 0.00140952 0.00210231
c hard coral boulder 0.13574759 0.13594954
c hard coral branching 0.09779898 0.09775753
c hard coral encrusting 0.07828255 0.07813595
c hard coral foliose 0.01474574 0.0143658
c hard coral mushroom 0.01843218 0.01857043
c hard coral submassive 0.01637211 0.01646811
c hard coral table 0.00173479 0.00175193
c soft coral 0.46882793 0.4688157
c soft coral gorgonian 0.0074813 0.00735809
c sponge 0.14008457 0.13980378
c sponge barrel 0.01149301 0.01156272
substrates.
Due to multiple difficulties that underwater images bear, strategies to en-
hance the image quality which should help to find better features are used. We
applied two of the state-of-the-art deep learning approaches and additionally
combined these with an improvement of our own development [3]. Those ap-
proaches are presented in the following subsections.
4.1 Image Enhancement
Underwater images inherit problems, like the attenuation of light or the suspen-
sion of particles reflecting the light. Those conditions distort colors and visibility,
which affect the performance of machine learning algorithms. Therefore we ap-
plied and evaluated multiple enhancement algorithms, that are specialized on
underwater images.
Ancuti et al. [1] use Fusion [24] and work without knowledge of a physical
model of the lighting conditions. Two derivations of the image, improving the
white balance and the contrast, are fused together using different weight mea-
sures to restore the image. The authors show that they retrieve more features
using SIFT [15] through applying their image enhancement. Ghani and Isa [7]
use Rayleigh-stretching, as well as stretching using the HSV color model, to first
correct the contrast and then correct the color. Further on the processings are
referred to as Fusion and RD. Since we do not have access to already corrected
images of the instant dataset, we use three evaluation measures that predict
how human would perceive the image, based on learned examples. Namely the
measures are BRISQUE [16], NIQE [17] and PIQUE [23]. For all of them smaller
values mean better image quality.
The evaluation [18] of the dataset of 2019 showed an improvement in image
quality using the two enhancement algorithms, as seen in table 4 as well as in
the subsections 4.4 and 5. Example images are shown in figure 1.
Table 4: Image enhancement algorithms evaluated on ImageCLEF Coral
dataset 2019.
Algorithm BRISQUE NIQE PIQUE
None 25.98 3.61 26.34
Fusion 22.66 3.43 30.99
RD 20.92 2.94 25.06
(a) Original (b) Fusion (c) RD
Fig. 1: Comparison of the image enhancements.
4.2 Yolo - Improvement
Neural networks are still state of the art in segmentation and classification tasks.
Last year we used Yolov3 [20] as our main neural network approach. Especially
areas with a particularly large denseness of smaller corals were challenging. The
reason for this is on one hand Yolo’s native ROI restraints which sets a natural
limit on the regions considered within a certain area. On the other hand we were
limited by the input data size with the largest resolution we could use with GPU
of 608 × 608 pixels. Regarding the original resolution of images of 4032 × 3024,
we preserve a scaling factor of at least of 5, i.e. each pixel represents an area of
more than 25 pixels in the original image. Considering that the smallest corals
consists of 12 × 12 pixels, the information loss is significant. We split the images
into overlapping subimages of 608 × 608 pixels and trained the network on
unscaled input. Consequently we got a large amount of images which also meant
that the training time of our network was almost one month. Unfortunately,
there was a mistake in the training data, therefore a revision of the results of the
last years challenge was just recently announced. This left us with not enough
time to retrain our working setup.
4.3 RetinaNET
For our second neural network approch we chose RetinaNet. The network showed
impressive results on the COCO dataset and outperformed Yolov2 with a 17%
higher AP0.5 value. RetinaNet is well suited, since it is able to produce more
predictions and is capable to work with less balanced data. At its core, the
architecture consists of the following components: a feature pyramid network
[12] (based on Resnet), a regressor for bounding box prediction and a classifier.
Basically, it is a one-stage detector. The particular advantage of RetinaNet is
the focal loss [14]. In case of end-to-end object detection, background predictions
oftentimes dominate. The optimizer rates the prediction as correct and the loss of
the positive background prediction forms the complete return loss. This mostly
leads to an optimizer return value of zero for the background areas in case of
cross entropy and thus reduce the loss. Focal loss weights the positive samples
higher and ensures that the network performs better on unbalanced data.
Right suited anchor boxes are the key to quality of object detection for any
architecture that works with a ”regions of interest”. If the anchors are not prop-
erly prepared, the network has in many cases no chance of finding particularly
small, neither large objects. In our dataset, we experience a wide variety of sizes
of corals. Starting from a box size of 18 × 9 to a size of 3966 × 2662 pixels,
the standard deviation of the areas is 629 assuming a square size. That means
that irregular or peripherally sized objects present a special challenge. To tackle
this problem, we chose a solution that was originally used on medical data [25].
In our opinion, the potential for improvement can be easily transferred to coral
context, since tumors and nodules as smaller objects are comparable to coral
objects of small size.
4.4 Own Developments
Besides the deep learning we also increased the performance of our classic feature
based approaches. We evaluated the use of principal component analysis (PCA)
[19] to select the best features, which increased the performance slightly. Apart
from the features the choice of the classifier is important. Last year we used
k-NN, which is depended from the parameter k. To overcome the search for
the right parameter we evaluated the use of naı̈ve bayes [21] for locating and
classifying substrates.
When classifying the coral areas and non-coral areas, the features along with
the approach are the same as in [3], which is illustrated in figure 2. A problem
that can be seen with k-NN is the low precision. This is due to labeling non-coral
areas as coral areas, because most often water gets falsely classified. There are
multiple ways to evaluate, as well as multiple things to evaluate. Beside the pure
evaluation of the coral and non-coral tiles, we also evaluate the bounding boxes
that enclose those coral tiles. In the end the evaluation of the found bounding
boxes is more significant, but the pure evaluation of the coral and non-coral tiles
helps with the assessment of the performance of finding bounding boxes in the
generated black and white images (see figure 2b - 2d). Likewise this creates the
opportunity to compare different image enhancement algorithms. All results are
discussed in the following.
Table 5 holds the evaluation of the coral/non-coral grid that is compared
with a grid representing the ground truth. The results show that the naı̈ve bayes
classifier has increased the accuracy, as well as the precision but greatly decreased
the recall. PCA on the other hand did not have a big impact.
It could be shown that image enhancement increases the performance, even
if just slightly. Table 6 showed that connected components works much better
with the naı̈ve bayes classifier that k-NN. The naı̈ve bayes classifier has an in-
significantly increased, whereas k-NN has significantly worse performance. We
believe that this is caused by k-NN having to much false-positives which results
in big boxes, that cover rather more than less area. Another indication for this
assumption is the decrease of precision but increase of recall (compare figure
2b).
Table 5: Evaluation of the coral/non-coral classification, based on the tiles.
Image enhancement
None Fusion RD
Approach Acc Prec Rec Acc Prec Rec Acc Prec Rec
k-NN 0.617 0.4496 0.5283 0.5823 0.4066 0.4813 0.6146 0.4355 0.4275
k-NN with PCA 0.6146 0.4471 0.5343 0.5858 0.4075 0.4637 0.6171 0.4371 0.4134
Bayes with PCA 0.6488 0.4545 0.1317 0.6570 0.3550 0.0031 0.6575 0.4857 0.0192
Table 6: Evaluation of the coral/non-coral classification, based on the bounding
boxes.
Image enhancement
None Fusion RD
Approach Acc Prec Rec Acc Prec Rec Acc Prec Rec
k-NN 0.4234 0.3920 0.8755 0.3919 0.4386 0.8543 0.4812 0.3973 0.8302
k-NN with PCA 0.4088 0.4088 0.8927 0.4097 0.4345 0.83106 0.4879 0.3888 0.8186
Bayes with PCA 0.6518 0.4833 0.2558 0.6571 0.3373 0.0025 0.6606 0.5948 0.0254
We also used the same two classifiers for classifying the bounding boxes.
While using the same set of features, k-NN outperformed naı̈ve bayes by far.
Image enhancement shows its contribution again with tripling the accuracy of the
naı̈ve bayes classier, as seen in table 7. This substantiate that the enhancement
proposed by Ghani and Isa [7] leads to increased performance with our dataset.
Table 7: Evaluation of substrate classification. The given values represent the
accuracy.
Image enhancement
Approach None Fusion RD
k-NN 0.4332 0.4550 0.4374
Bayes 0.1261 0.3328 0.3672
5 Evaluation of the Submitted runs
The evaluation was processed on test data that consists of images from four
different geographical regions than the training set:
– same location
– similar location
– geographically distinct but ecologically connected
– geographically and ecologically distinct
In total the test dataset has 400 images, made by 100 images per subset. The
results are examined in more detail below, while the interesting and informative
values are discussed in the text. For the complete list of results, the reader
is referred to the task overview working note [6]. Each submitted result was
produced by one of our approaches, all of which were trained on the full training
set.
The classification of the substrate types based on the classic features alone
fails largely with an MAP0 value of around 27.4. and an MAP0.5 of 1. Although
we improved the MAP0.5 value compared to last year by 300 percent, the results
are still not really useful due to the low absolute values. This applies to both
experiments, for the classification by means of k-NN based on the chosen features
boosted by PCA and to statistical label assignment as well.
Our neuronal Network based approaches show a significantly better perfor-
mance.
Since we had a limitation in the number of submissions, we chose none lin-
ear composition of available options. Our pool of options consisted alongside to
classical features of RetinaNet and Yolov3, which both were trained on the un-
preprocessed data and also on enhanced images by RD and Fusion. In addition,
we worked with a variation of threshold τ ∈ {0.001, 0.1, 0.2, 0.5}, which limits
the MAP.
(a) Grund truth boxes. (b) Inside (white) and outside tiles (black)
using k-NN.
(c) Inside (white) and outside tiles (black) (d) Inside (white) and outside tiles (black)
using PCA and k-NN. with PCA and naive Bayes.
Fig. 2: Visualized process of localization corals with our approach. The raw pic-
ture is taken from the ImageCLEFcoral dataset [6].
We achieve our best result with an ensemble of RetinaNet and Yolov3.
Whereby the predictions of RetinaNet were extended by the predictions of
Yolov3. Both Networks were trained on RD-preprocessed images, with a τ of
0.1. The combination of both systems results in a MAP0.5 of 39.2 % and MAP0
of 80.6 %. It is noticeable that we predicted very few bounding boxes for a
MAP0.5 both through Yolov3 and through RetinaNet, additionally the found
predictions were far from being present in all of the images.
In case of reduction of the accepted overlap, we get significantly more bound-
ing boxes which also leads to a higher chance of hitting the right coral within the
test images on cost of our overall accuracy. However, if we increase the accuracy,
the MAP value drops. The following are significant examples: RetinaNet (τ =
0.01) combined with Yolov3 (τ = 0.01) has an MAP0.5 of 0.303 and an MAP0
of 0.727, but only an overall accuracy of 7 %. In contrast to it RetinaNet alone
produces significantly fewer boxes with (τ = 0.2), but achieves the best overall
accuracy of 14.2 % with an MAP0.5 of only 30.3 and a MAP0 of 66.3 % while
the accuracy is 10 % higher than in our best run.
A possible explanation is probably that RetinaNet has not finished training
in 35 epochs. The large amount of predictions with a low overlap value could
probably be boosted by non-maximum suppression. However, we were not aware
of this problem until the results were published.
When provided with such a variety of possibilities, it can not be clearly deter-
mined which enhancement variant is the best choice. Certainly enhanced images
improve the predictions, while we suspect that RD is superior to Fusion for coral
images.
5.1 Transferability of the results within the test data subsets
One major question is the transferability of the results within the test dataset,
considering that results, that form the average measure, vary widely. For our
best run the MAP0.5 for same location increases by 6 percent to 45.7 % and by
1 % to 81.5 % the MAP0 . For similar location, however, it drops to 28.3 % for
MAP0.5 , but reaches a value of 86.4 % for MAP0 and thus has the highest score
among all presented submissions by all participants. Our best approach performs
very well on geographically similar data. The MAP0.5 is 42.6 % and the MAP0
81.2 %. Overall, this part of the test dataset seems to be less complex, which is
shown by the fact that the performance of almost all of our approaches have an
increased performance on it.
It is also worth mentioning that we perform significantly worse on geographi-
cally distinct data with just an MAP0.5 0.125 and an MAP0 0.362. This tendency
is also evident in the other approaches we used. The data seems to differ signif-
icantly, we either generalize less or the classes that are harder to recognize are
more present. Some additional, yet unknown substrate types may be more dom-
inant in geographically and ecologically distinct rocky reef and lead to distorted
results. A further investigation of the dataset is required.
6 Conclusion
Overall, our approaches show significantly better results than last year. A com-
parison of our best approaches between the two years shows that we have im-
proved the MAP0.5 by 13.4 %. The classification of the tiles into in- and outside
boxes could also be improved by a Bayesian classifier, but is still far from being
accurate.
Image enhancement techniques on the last years data were confirmed by the
evaluation of the current test dataset, which leads us to the conclusion that the
correction of blurry images in terms of contrast and sharpness is necessary. The
RD algorithm makes the greatest contribution to improving the quality of the
images. Less noteworthy results are made with the classical feature engineering
approach. A deeper examination of the features and their information value is
needed.
It can be assumed that the fair partitioning of images according to our bal-
ancing strategy, also has significantly contributed to the improved results. Deep
learning strategies generalize quiet well and are superior when using for this
task. Especially the performance of RetinaNet, since it is not only better on
the coco dataset than the state of the art. So far, the complexity of the images
can hardly be handled by a single approach. We still see the most potential in
an ensemble of several architectures. The combination of advantages of different
approaches is the key to a stable solution. Since the amount of data has grown,
while remaining relatively small, we still cannot exclude the potential of classic
machine learning. Usually neural networks show a much better performance with
an increased amount of data. For this reason, the images should be split into
overlapping images and thus increase the number of training samples.
For future approaches, we recommend the usage of more specific features
that are suitable for corals, such as used in [2]. However, we would rather rely
on an ensemble of neural networks of the gold standard, which we would reduce
in depth to shorten the training time and increase in terms of input resolution
to decrease the information loss.
References
1. Ancuti, C., Ancuti, C.O., Haber, T., Bekaert, P.: Enhancing underwater images
and videos by fusion. In: 2012 IEEE Conference on Computer Vision and Pattern
Recognition. pp. 81–88. IEEE (2012)
2. Ani Brown Mary, N., Dharma, D.: Coral reef image classification employ-
ing improved ldp for feature extraction. J. Vis. Comun. Image Repre-
sent. 49(C), 225–242 (Nov 2017). https://doi.org/10.1016/j.jvcir.2017.09.008,
https://doi.org/10.1016/j.jvcir.2017.09.008
3. Bogomasov, K., Grawe, P., Conrad, S.: A two-staged approach for localization
and classification of coral reef structures and compositions. In: Working Notes of
CLEF 2019 - Conference and Labs of the Evaluation Forum, Lugano, Switzerland,
September 9-12, 2019 (2019), http://ceur-ws.org/Vol-2380/paper 106.pdf
4. Caridade, C.M.R., Marçal, A.R.S.: Automatic classification of coral images using
color and textures. In: CLEF (2019)
5. Chamberlain, J., Campello, A., Wright, J.P., Clift, L.G., Clark, A., Garcı́a Seco de
Herrera, A.: Overview of ImageCLEFcoral 2019 task. In: CLEF2019 Working
Notes. CEUR Workshop Proceedings, CEUR-WS.org (2019)
6. Chamberlain, J., Campello, A., Wright, J.P., Clift, L.G., Clark, A., Garcı́a Seco de
Herrera, A.: Overview of the ImageCLEFcoral 2020 task: Automated coral reef
image annotation. In: CLEF2020 Working Notes. CEUR Workshop Proceedings,
CEUR-WS.org (2020)
7. Ghani, A.S.A., Isa, N.A.M.: Underwater image quality enhancement through com-
position of dual-intensity images and rayleigh-stretching. SpringerPlus 3(1), 757
(2014)
8. Hoegh-Guldberg, O., Mumby, P.J., Hooten, A.J., Steneck, R.S., Greenfield, P.,
Gomez, E., Harvell, C.D., Sale, P.F., Edwards, A.J., Caldeira, K., et al.: Coral
reefs under rapid climate change and ocean acidification. science 318(5857), 1737–
1742 (2007)
9. Ionescu, B., Müller, H., Péteri, R., Abacha, A.B., Datla, V., Hasan, S.A., Demner-
Fushman, D., Kozlovski, S., Liauchuk, V., Cid, Y.D., Kovalev, V., Pelka, O.,
Friedrich, C.M., de Herrera, A.G.S., Ninh, V.T., Le, T.K., Zhou, L., Piras, L.,
Riegler, M., l Halvorsen, P., Tran, M.T., Lux, M., Gurrin, C., Dang-Nguyen, D.T.,
Chamberlain, J., Clark, A., Campello, A., Fichou, D., Berari, R., Brie, P., Dogariu,
M., Ştefan, L.D., Constantin, M.G.: Overview of the imageclef 2020: Multimedia
retrieval in medical, lifelogging, nature, and internet applications. In: Experimental
IR Meets Multilinguality, Multimodality, and Interaction. Proceedings of the 11th
International Conference of the CLEF Association (CLEF 2020), vol. 12260. LNCS
Lecture Notes in Computer Science, Springer, Thessaloniki, Greece (September 22-
25 2020)
10. Jaisakthi, S.M., Mirunalini, P., Aravindan, C.: Coral reef annotation and localiza-
tion using faster r-cnn. In: CLEF (2019)
11. Lin, J.: Divergence measures based on the shannon entropy. IEEE Transactions on
Information theory 37(1), 145–151 (1991)
12. Lin, T., Dollár, P., Girshick, R.B., He, K., Hariharan, B., Belongie, S.J.: Fea-
ture pyramid networks for object detection. CoRR abs/1612.03144 (2016),
http://arxiv.org/abs/1612.03144
13. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object
detection. In: Proceedings of the IEEE international conference on computer vision.
pp. 2980–2988 (2017)
14. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for
dense object detection. 2017 IEEE International Conference on Com-
puter Vision (ICCV) (Oct 2017). https://doi.org/10.1109/iccv.2017.324,
http://dx.doi.org/10.1109/ICCV.2017.324
15. Lowe, D.G.: Distinctive image features from scale-invariant keypoints. Interna-
tional journal of computer vision 60(2), 91–110 (2004)
16. Mittal, A., Moorthy, A.K., Bovik, A.C.: No-reference image quality assessment in
the spatial domain. IEEE Transactions on image processing 21(12), 4695–4708
(2012)
17. Mittal, A., Soundararajan, R., Bovik, A.C.: Making a “completely blind” image
quality analyzer. IEEE Signal processing letters 20(3), 209–212 (2012)
18. Nguyen, H.C.: Comparison of methods for underwater image improvement (2020)
19. Pearson, K.: Liii. on lines and planes of closest fit to systems of points in space. The
London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science
2(11), 559–572 (1901)
20. Redmon, J., Farhadi, A.: Yolov3: An incremental improvement. arXiv preprint
arXiv:1804.02767 (2018)
21. Rish, I., et al.: An empirical study of the naive bayes classifier. In: IJCAI 2001
workshop on empirical methods in artificial intelligence. vol. 3, pp. 41–46 (2001)
22. Steffens, A., de A. Campello, A.C., Ravenscroft, J., Clark, A., Hagras, H.: Deep
segmentation: using deep convolutional networks for coral reef pixel-wise parsing.
In: CLEF (2019)
23. Venkatanath, N., Praneeth, D., Bh, M.C., Channappayya, S.S., Medasani, S.S.:
Blind image quality evaluation using perception based features. In: 2015 Twenty
First National Conference on Communications (NCC). pp. 1–6. IEEE (2015)
24. Xydeas, C., , Petrovic, V.: Objective image fusion performance measure. Electron-
ics letters 36(4), 308–309 (2000)
25. Zlocha, M., Dou, Q., Glocker, B.: Improving retinanet for ct lesion detection with
dense masks from weak recist labels. In: International Conference on Medical Image
Computing and Computer-Assisted Intervention. pp. 402–410. Springer (2019)