Similar image retrieval using convolutional neural networks: A study of feature extraction techniques

• Main Authors: Mtasher, Ashwaq Katham, Msad, Jenan Jader
• Format: Conference Proceeding
• Language: English
• Subjects: Accuracy; Algorithms; Artificial neural networks; Datasets; Euclidean geometry; Feature extraction; Image analysis; Image detection; Image enhancement; Image retrieval; Machine learning; Neural networks; Pictures; Representations; Retrieval; Searching; Similarity
• ISSN: 0094-243X; 1551-7616
• DOI: 10.1063/5.0199703

This work presents a machine learning approach for detecting related photos. Using a convolutional neural network (CNN), our system extracts features from pictures and compares the feature vectors using a similarity metric. We test our algorithm on a massive dataset of photographs and demonstrate that it can efficiently and accurately discover related images. We also compare our approach to established techniques such as SIFT and SURF, showing that it outperforms them in terms of accuracy and computing economy. The suggested approach has applications in image search, duplicate picture identification, and image retrieval systems. The proposed machine learning approach for detecting comparable photos uses a convolutional neural network (CNN) to extract features from images and a similarity measure to compare feature vectors. A large dataset of pictures is used to train the CNN to create a feature representation that captures the semantic and visual features of the images. Once introduced, the CNN can extract features from new photos and compare them to components from other images to find similar images. One of the primary benefits of utilizing a CNN for image feature extraction is its ability to learn a hierarchical representation of the pictures, allowing it to collect both low-level and high-level information. In contrast, existing methods such as SIFT and SURF often capture only low-level characteristics such as edges and corners. A similarity metric is used to compare the feature vectors. Cosine similarity, Euclidean distance, or any other similarity metric that can represent the similarity between two vectors may be used.The results of our study demonstrate the effectiveness and efficiency of the suggested technique for discovering similar images. Extensive testing on a large dataset of photographs confirms its ability to efficiently identify comparable pictures. In comparison to traditional approaches like SIFT and SURF, our algorithm outperforms them in both accuracy and computational efficiency.The implications of our findings extend to various applications in the field of image analysis. Specifically, our approach can greatly enhance image search, duplicate picture identification, and image retrieval systems. In the context of image search, it enables the identification of similar images to a given query image. Additionally, in duplicate image detection, it can efficiently identify identical photos within a collection.Moreover, our technique proves