Page 158 - 2023-Vol19-Issue2
P. 158
154 | Mohammed, Oraibi & Hussain
Fig. 8. MobileNet Validation.
Fig. 7. Ensemble Learning CM.
model is able to identify the classes with a high degree of Fig. 9. Inception Validation.
ACC. The majority of the classes have a PREC, REC and
F1-S of at least 0.90, with some classes having a perfect score which have the less distance , the same with mobile phone the
of 1.00. The ACC is good but the PREC, REC and F1-S of picture number 6 is the closest to the query image.
some classes can be improved.
B. Validation Results C. Comparison with State-of-the-Art Methods
In order to validate our proposed methods for image similarity, In this study, we evaluated four different approaches for image
we implemented a program that would output the top 10 most classification: Xception, Inception, MobileNet, and Ensemble
similar images for a given input image. The program first pre- Learning (EL). The results of our experiments show that all
processed the images by resizing them to a standard resolution four methods have high ACC, PREC, REC, and F1-S. How-
and converting them to grayscale. Next, the program extracted ever, when we compare our approaches to other papers in the
features from the images using a pre-trained DL model. These literature, we found that our Ensemble Learning approach
features were then used to calculate the similarity between outperforms the others, with a 95% ACC, PREC, REC and
the input image and all the other images in the dataset using a F1-S. This highlights the effectiveness of our EL method in
distance metric such as cosine similarity or Euclidean distance. image classification tasks, and suggests that it could be a valu-
After calculating the similarity scores, the program sorted the able tool in various applications. Additionally, our results
images in descending order of similarity and selected the top also indicate that Xception, Inception, and MobileNet are all
10 most similar images. These images were then displayed to strong contenders, achieving similar high performance. Our
the user along with their similarity scores, allowing the user to experiments demonstrate the robustness and effectiveness of
easily compare and evaluate the similarity between the input these four approaches in image classification as in Table I.
image and the top 10 most similar images. It is worthy to mention that we only applied our Hard Vot-
ing ensemble approach on 20 classes out of 50 provided by
Additionally, the program also allowed users to experi- CBIR 50 dataset. This is because of the limitations imposed
ment with different pre-trained models and distance metrics to by Google Colaboratory. In the future we will conduct the
see how these variations affected the similarity scores and the experiments on the full dataset.
final selection of the top 10 most similar images. The program
was able to process large datasets of images in a relatively The choice of Deep Learning models in our study is
short amount of time, making it an efficient and practical tool guided by their unique capabilities and proven performance
for image similarity evaluation. We validate it in those two in image classification tasks. We employ the Xception model,
methods ”MobileNet and Inception” as Figure 8 and Figure 9 an advanced form of the Inception architecture, which uses a
shown, for the camera the closer one is the picture number 3 linear stack of depth-wise separable convolution layers with
