Comparison of AlexNet Algorithm with DenseNet Algorithm for

Aquatic Debris Detection on Ocean Surfaces

T. M. Kamal and N. Bharatha Devi

Department of Computer Science and Engineering, Saveetha School of Engineering, Saveetha Institute of Medical and

Technical Sciences, Chennai, Tamil Nadu, India

Keywords: Novel AlexNet Algorithm, Convolutional Neural Network, Deep Learning, DenseNet Algorithm, Marine

Debris, Ocean Surfaces, Research.

Abstract: This research aims to spot aquatic debris on ocean surfaces using the novel Alexnet algorithm, comparing its

performance against the DenseNet algorithm. Materials and Methods: To enhance the detection of aquatic

debris, two algorithms such as Novel AlexNet and DenseNet are compared with iteration done for each group

as 20 are implemented on the data set with a G-Power of 80% and confidence level of 95% using the clinical

software. The dataset used for this research consists of 324 JPG images of marine debris and 724 JPG images

of the ocean with varying degrees of clarity, including images with noise. Results: With an accuracy of

93.77%, the Novel AlexNet algorithm identifies and measures objects with more accuracy than the DenseNet

algorithm, with 92.91%. It shows that there is no statistical significance difference between the Novel Alexnet

Algorithm and DenseNet Algorithm with p=0.530 (Independent Sample T-test p<0.05). Conclusion: The

detection of aquatic debris using the Novel AlexNet algorithm provides superior performance compared to

the DenseNet algorithm, as concluded from the obtained results.

1 INTRODUCTION

The escalating volume of waste in the Earth's oceans

has emerged as a pressing environmental issue,

posing risks to ecosystems, wildlife, and human well-

being. Debris finds its way into the oceans through

rivers, sewage outlets, and wind transport.

Furthermore, maritime vessels contribute to the

ocean's waste accumulation, resulting in the

formation of garbage patches (Politikos, 2021).

Ocean currents and winds occasionally transport

marine debris great distances from its origin in the

ocean. Within the sea, specific regions known as

garbage patches emerge as locations where marine

debris gathers due to prevailing currents. Machine

learning, a recent field in Computer Science,

encompasses a range of data analysis methods (Chen,

2022). A portion of these methods relies on

established statistical approaches like logistic

regression and principal component analysis, while

several others do not. Deep learning gained

momentum in object recognition and detection

through the introduction of AlexNet (Huang, 2021).

A group of scientists employed deep learning in their

investigation of image-based water recognition, and

over time, the methodology advanced in complexity

(Y. Wang, 2022) (James et al., 2017). The application

of deep learning techniques and their applications can

increase the efficiency of the process by automating

the removal of marine garbage, discovering its

distribution, and improving debris identification.

Most statistical techniques aim to identify the

optimal probabilistic model describing observed data

within a related model class. Similarly, in machine

learning, the focus is on finding models that best suit

the data, solving optimization problems. Unlike

before, these machine learning models aren't confined

to probabilistic models. (Marin et al., 2021), (G.

Ramkuamr et al., 2021). These regions are visible to

the naked eye and consist predominantly of

microplastics.

2 LITERATURE SURVEY

In the last 4 years, there have been 81 articles in IEEE

Explorer and 2250 articles in Google Scholar. Many

publications that focus on various applications were

produced. There are a few thorough studies on the

identification of deep-sea debris, however, the

Kamal, T. and Devi, N.

Comparison of AlexNet Algorithm with DenseNet Algorithm for Aquatic Debris Detection on Ocean Surfaces.

DOI: 10.5220/0012772400003739

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

In Proceedings of the 1st International Conference on Artiﬁcial Intelligence for Internet of Things: Accelerating Innovation in Industry and Consumer Electronics (AI4IoT 2023), pages 367-372

ISBN: 978-989-758-661-3

367

majority of marine debris detection exclusively

concentrates on the sea surface and coast (Vethaak,

2022). Although certain detection networks have

started to be deployed to find trash in the ocean, the

results have not been adequate(Wolf et al. 2021). In

the prior research titled "A Deep Feature-Based

Strategy for Categorizing Marine Debris," the

suggestion is put forth to employ the DenseNet

Algorithm for recognizing and categorizing marine

debris. The study assesses how a neural network

classifier, trained using deep CNN feature extractors,

performs when the feature extractor is held constant

versus when it's adapted in the given task (Huang

2021) (Sivakumar et al 2022). Annotations at the

image level, indicating the existence or absence of

target objects, prove adequate for Weakly Supervised

Learning (Y. Wang 2022). Weakly Supervised

Learning merely necessitates image-level annotations

that specify whether the target objects are present in

an image or not. In this study, the AlexNet algorithm

is employed as the backbone network for object

classification. The limitation of this study involves

the replacement of the final two fully connected

layers with two convolutional layers and a Global

Average Pooling layer. In contrast, the current paper

employs a configuration of five convolutional layers

and eight interconnected layers (Wu, 2020).

The research gap was identified using the present

technique with poor accuracy. The existing method

has shortcomings, such as the significant amount of

information required for accurate forecasting. While

using less data for testing and training, the

recommended solution, the Novel AlexNet

algorithm, delivers improved accuracy. The present

research tests the detection of aquatic debris on ocean

surfaces using the Novel Alexnet algorithm and the

DenseNet algorithm.

3 PROPOSED METHODOLOGY

Image Processing Lab is utilized to conduct the

research in the Department of Computer Science and

Engineering, Saveetha School of Engineering,

Saveetha Institute of Medical and Technical Sciences,

Chennai. The study involved selecting two groups

and comparing their processes to derive results

(Freitas, Silva, and Silva, 2021). G power value is

obtained as 80% where alpha value is 0.05 and 0.2 is

beta value with a 95% confidence interval. A sample

is determined to 20 for each group

(https://clincalc.com/stats/samplesize.aspx).

Utilizing the MATLAB software, the proposed

work was designed and implemented, with the dataset

for the code implementation being provided by

Kaggle. This dataset comprises 724 jpg file format

images of clear ocean and 324 jpg file format images

of aquatic debris, allowing for the analysis of both

clear ocean and debris records (Y. Wang, 2022). The

Windows 11 OS served as an experimental

environment for deep learning. A minimum of 4 GB

of RAM was used in the hardware setup, which

included an Intel Core i3 CPU. 64-bit system sort was

employed. Utilizing MatLab code, Code is

implemented. To execute an output process for

accuracy during code execution, the image set is

processed behind this code.

3.1 AlexNet Algorithm

Novel AlexNet Algorithm is used as a sample

preparation of group 1. Eight layers comprise the

Novel AlexNet algorithm, including three

interconnected and five convolutional layers.

Convolutional filters and ReLU, a nonlinear

activation function, are both included in every

layer(Kang et al., 2022). Rectified Linear Units

(ReLU) are used by Novel AlexNet to help models

learn and solve issues more quickly, as well as

address difficulties with disappearing gradients. The

first convolutional network used to improve GPU

performance is called Novel AlexNet (Alom et al.,

2021). The results of Novel AlexNet show that a

large-scale, deep convolutional neural network can

break records on a highly challenging dataset with

just supervised learning. A new phase of research was

inaugurated by CNN's Novel AlexNet, the company's

inventor. The implementation of Novel AlexNet is

rather straightforward given the abundance of deep

learning frameworks available (Wu, 2021).

3.2 Procedure of Novel AlexNet

Algorithm

Input - Dataset records

1. Apply pre-processing techniques to the object

images to eliminate any background noise.

2. Use the threshold algorithm to convert the pre-

processed image to a binary image.

3. Identification of the parts of the binary images must

be carried out.

4. Segment the individual object from the image.

5. Extract the major axis, minor axis, and area

geometric features of each individual object.

6. Evaluate the quality by analyzing the average

feature values extracted from the sample.

7. Classify the sample according to its type and grade,

based on the analysis performed.

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

368

3.3 DenseNet Algorithm

In the context of preparing samples for group 2, the

DenseNet algorithm is employed. DenseNet operates

by combining the output of the preceding layer with

that of the following layer, while ResNet employs an

additive approach (+) to merge the previous layer

(identity) with the subsequent layer (Kikaki et al.,

2022). The main purpose behind developing

DenseNet was to address the reduced accuracy caused

by the vanishing gradient problem in deep neural

networks. To put it simply, the information vanishes

along the way due to the extensive path between the

input and output layers (G. Wang et al., 2021).

Through the utilization of the composite function

operation, the output from the initial layer is used as

input for the subsequent one. This composite

procedure involves components like convolution

layers, pooling layers, batch normalization, and non-

linear activation (Li and Wang, 2022). The network

has L(L+1)/2 direct connections as a result of this

linkage. L is the number of architectural layers. There

are many variations of the DenseNet, including

DenseNet-121, DenseNet-160, and DenseNet-201.

The numerical values represent the neural network's

layer count.

3.4 Procedure for DenseNet Algorithm

Input - Training Set

1. Load the pre-trained DenseNet model and create

a new model by adding a global average pooling layer

and a fully connected layer with sigmoid activation

for binary classification.

2. Compile the new model using an optimizer, loss

function, and evaluation metrics.

3. Augment the training data using the

ImageDataGenerator function to reduce overfitting.

4. Train the new model using the fit_generator

method and evaluate it on the validation set using the

evaluate_generator method.

5. Fine-tune the DenseNet model by unfreezing

some of its layers and retraining on the augmented

data with a smaller learning rate. Use the trained

model to predict the presence of aquatic debris in

ocean surface images.

Statistical Analysis

To perform statistical analysis of Novel AlexNet and

GoogleNetDenseNet, SPSS software was employed.

The dependent variables consisted of image

characteristics such as objects, distance, frequency,

modulation, pixel distribution, while the independent

variables included frame rate and resolution. The

statistical software used for analysis was IBM SPSS

version 26. For both algorithms, datasets were

generated using a sample size of 10. Testing variables

for accuracy and loss were employed, and grouping

was performed using Group ID.

4 RESULTS

The effectiveness of the Novel AlexNet algorithm in

predicting aquatic debris in ocean surfaces was

compared to that of the ResNet-50 algorithm and the

accuracies are found to be 93.77% and 92.91%

respectively.

Novel AlexNet and DenseNet's mean and

accuracy results are shown in Table 1. Novel

AlexNet's mean value is better than DenseNet's,

which has a standard deviation of 1.63330 and

1.41324, respectively.

The results of Novel AlexNet and DenseNet's

independent sample T-test are displayed in Table 2. It

demonstrates no statistically significant difference

between the Novel Alexnet Algorithm and the

DenseNet Algorithm with a value of p=0.530

(Independent Sample T-test p0.05).

Figure 1 illustrates the mean accuracy and loss

compared with Novel AlexNet and DenseNet. The

Novel AlexNet and ResNet T-test results are

displayed as a bar graph. It displays the Novel

AlexNet and DenseNet accuracy and loss numbers.

The accuracy numbers looked like bars.

Table 1: Group Statistics between Novel AlexNet algorithm and DenseNet algorithm.

Group

Mean

Std. Deviation

Std. Error Mean

Accuracy

AlexNet

93.7730

1.63330

0.36522

DenseNet

92.9125

1.41324

0.31601

Loss

AlexNet

4.2160

1.65847

0.37085

DenseNet

7.8355

1.47372

0.32953

Comparison of AlexNet Algorithm with DenseNet Algorithm for Aquatic Debris Detection on Ocean Surfaces

369

Table 2: Independent Sample T-Test. It shows that there is no statistical significance difference between the Novel Alexnet

Algorithm and DenseNet Algorithm with p=0.530 (Independent Sample T-test p<0.05).

Leven’s Test of Equality

of Variances

T-test for Equality of Means

95% Confidence Interval

of the Difference

Sig

(2-tailed)

Mean

Difference

Std Error

difference

Lower

Upper

Accuracy

Equal

Variance

assumed

3.347

0.075

0.634

.530

.30600

.48296

-671

1.283

Equal

Variance not

assumed

0.634

37.231

.530

.30600

.48296

-672

1.2843

Loss

Equal

Variance

assumed

3.814

0.058

-1.971

.055

-.97800

.49610

-1.982

.2631

Equal

Variance, not

assumed

-1.971

37.482

.056

-.97800

.49610

-1.98277

.2677

Figure 1: Comparison of Novel AlexNet and DenseNet in terms of mean accuracy and loss. The mean accuracy of Novel

AlexNet is better than DenseNet. X-axis: Novel AlexNet VS DenseNet, Y-axis: Mean accuracy: Error bar with +/- 2SD.

5 DISCUSSION

The study's findings revealed that the significance

value obtained is 0.011 (p<0.05), indicating a

statistical significant difference between the two

groups. Moreover, based on the accuracy analysis, the

Novel AlexNet algorithm was found to perform better

than DenseNet. Specifically, the Novel AlexNet

classifier achieved an accuracy of 93.77%, whereas

the GoogleNetDenseNet classifier had an accuracy of

92.40%.

The research work (Marin et al., 2021) examined

multiple approaches for extending the connection of

a CNN in the time domain to determine the image and

enhance the feature of debris. It achieves an accuracy

of 91 % using DenseNet in the research.

Classification of deep learning CNN is said to be a

type of artificial neural network that detects and

recognizes objects using the previously saved image

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

370

data set (Freitas, Silva, and Silva, 2021). The main

goal of pattern recognition is to accurately classify an

input pattern as one of the several output classes. The

two main steps of this approach are feature selection

and classification (Politikos, 2021). Since the

classifier won't be able to distinguish between badly

chosen features, feature selection is essential to the

entire process (Kikaki et al., 2022). This research

application can be used in classifying the Debris on

Ocean surfaces and shores. A few other applications

include with the help of this approach the amount of

garbage that is mixed with water bodies is determined

and also used to classify the debris in the ocean

(Vethaak, 2022).

The factors affecting the study include the

collection of datasets from various sources. Another

factor that affects the study is the time required to

classify the data for both classification and detection.

The limitation of this study, the debris present below

the surface cannot be determined due to the lack of

light below the surface. The future scope of this study

includes the way to introduce hybrid algorithms to

improve accuracy. The future scope of this study is

that the system should be expanded to include a larger

number of objects with lesser time consumption in

training the data set.

6 CONCLUSION

The Aquatic Debris Recognition Data set was used to

improve the performance of image classification. The

study employed the Novel AlexNet algorithm and

DenseNet for this purpose. The mean accuracy of the

Novel AlexNet algorithm was found to be 93.77%,

whereas that of DenseNet was 92.91%. It can be

inferred that the Novel AlexNet algorithm yields

higher accuracy than the DenseNet algorithm.

REFERENCES

Alom, Md Zahangir, Tarek M. Taha, Christopher Yakopcic,

Stefan Westberg, Paheding Sidike, Mst Shamima

Nasrin, Brian C. Van Esesn, Abdul A. S. Awwal, and

Vijayan K. Asari, 2021. “The History Began from

AlexNet: A Comprehensive Survey on Deep Learning

Approaches.” http://arxiv.org/abs/1803.01164.

Chen, Ge., 2022. “Deep-Sea Debris Identification Using

Deep Convolutional Neural Networks.” 2022.

https://doi.org/10.1109/JSTARS.2021.3107853.

Freitas, Sara, Hugo Silva, and Eduardo Silva, 2021.

“Remote Hyperspectral Imaging Acquisition and

Characterization for Marine Litter Detection.” Remote

Sensing 13 (13): 2536.

G. Ramkumar, R. Thandaiah Prabu, Ngangbam Phalguni

Singh, U. Maheswaran, 2021, Experimental analysis of

brain tumor detection system using Machine learning

approach, Materials Today: Proceedings, ISSN 2214-

7853, https://doi.org/10.1016/j.matpr.2021.01.246.

Huang, Baoxiang. 2021. “An Efficient Deep-Sea Debris

Detection Method Using Deep Neural Networks”,

2021. https://doi.org/10.1109/JSTARS.2021.3130238.

James, J., Lakshmi, S. V., & Pandian, P. K., 2017. A

preliminary investigation on the geotechnical properties

of blended solid wastes as synthetic fill material.

International Journal of Technology, 8(3), 466-476.

Kang, Kai, Hongsheng Li, Junjie Yan, Xingyu Zeng, Bin

Yang, Tong Xiao, Cong Zhang, et al., 2022. “T-CNN:

Tubelets with Convolutional Neural Networks for

Object Detection from Videos.” IEEE Transactions on

Circuits and Systems for Video Technology 28 (10):

2896–2907.

Kishore Kumar, M. Aeri, A. Grover, J. Agarwal, P. Kumar,

and T. Raghu, “Secured supply chain management

system for fisheries through IoT,” Meas. Sensors, vol.

25, no. August 2022, p. 100632, 2023, doi:

10.1016/j.measen.2022.100632.

Kikaki, Katerina, Ioannis Kakogeorgiou, Paraskevi Mikeli,

Dionysios E. Raitsos, and Konstantinos Karantzalos,

2022. “MARIDA: A Benchmark for Marine Debris

Detection from Sentinel-2 Remote Sensing Data.” PloS

One 17 (1): e0262247.

Li, Ya-Li, and Shengjin Wang, 2022. “BooDet: Gradient

Boosting Object Detection with Additive Learning-

Based Prediction Aggregation.” IEEE Transactions on

Image Processing: A Publication of the IEEE Signal

Processing Society 31 (March): 2620–32.

Marin, Ivana, Saša Mladenović, Sven Gotovac, and Goran

Zaharija. 2021. “Deep-Feature-Based Approach to

Marine Debris Classification.” NATO Advanced

Science Institutes Series E: Applied Sciences 11 (12):

5644.

Politikos, Dimitris V., 2021. “Automatic Detection of

Seafloor Marine Litter Using Towed Camera Images

and Deep Learning.” Marine Pollution Bulletin 164

(March): 111974.

Sivakumar, V. L., Nallanathel, M., Ramalakshmi, M., &

Golla, V., 2022. Optimal route selection for the

transmission of natural gas through pipelines in

Tiruchengode Taluk using GIS–a preliminary study.

Materials Today: Proceedings, 50, 576-581.

Vethaak, Dick, 2022. “Marine Microplastic Debris: An

Emerging Issue for Food Security, Food Safety and

Human Health.” Marine Pollution Bulletin 133

(August): 336–48.

V. P. Parandhaman, "An Automated Efficient and Robust

Scheme in Payment Protocol Using the Internet of

Things," (2023) Eighth International Conference on

Science Technology Engineering and Mathematics

(ICONSTEM), Chennai, India, 2023, pp. 1-5, doi:

10.1109/ICONSTEM56934.2023.10142797.

Wang, Gaihua, Zhao Guo, Xizhou Wan, and Xu Zheng.

2021. “Study on Image Classification Algorithm Based

Comparison of AlexNet Algorithm with DenseNet Algorithm for Aquatic Debris Detection on Ocean Surfaces

371

on Improved DenseNet.” Journal of Physics.

Conference Series 1952 (2): 022011.

Wang, Yawen, 2022. “A Literature Review of Underwater

Image Detection.” Design Studies and Intelligence

Engineering L.C. Jain et Al. (Eds.) 1: 10.

Wolf, Mattis, Katelijn van den Berg, Shungudzemwoyo P.

Garaba, Nina Gnann, Klaus Sattler, Frederic Stahl, and

Oliver Zielinski, 2021. “Machine Learning for Aquatic

Plastic Litter Detection, Classification and

Quantification (APLASTIC-Q).” Environmental

Research Letters: ERL [Web Site] 15 (11): 114042.

Wu, Zhi-Ze, 2020. “Convolutional Neural Network Based

Weakly Supervised Learning for Aircraft Detection

From Remote Sensing Image.” 2020.

https://doi.org/10.1109/ACCESS.2020.3019956.

———. 2021. “Convolutional Neural Network Based

Weakly Supervised Learning for Aircraft Detection

From Remote Sensing Image.” 2021.

https://doi.org/10.1109/ACCESS.2020.3019956.

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

372