Content-Based Image Retrieval using Multi Deep Neural Networks and K-Nearest Neighbor Approaches

Abstract

With the rapid development of digital image technology, the number of digital images stored on the Internet environment has been increasing remarkably over the past decade. As a result, it has become a top priority to come up with an effective and convenient search tools for images. Although several image-based search tools have been introduced till now to allow users to search for images with relatively fast response times, it remains certain limitations from these tools in resolving ambiguity between content of the query image and returned image. In this research, we proposed a method exploiting multi deep neural networks and K-Nearest Neighbor methodologies in Content Based Image Retrieval to improve the image search quality. First, we built our own multi CNN models inherited from the pre-trained CNN models to extract image features. Next, basing on the method of K-Nearest Neighbors, we compiled the similarity measure of the distance between the feature vectors and conducted model experiments. Experiments were performed on the Oxford-IIIT Pet Image Dataset and self-collected dataset of a Kaggle competition. We then evaluated the model by accuracy, confusion matrix, and F1 scores; average precision (AP) and mean average precision (mAP) were used to evaluate the search results of our search proposed system. The experiments show that our proposed system achieved outstanding results.

An Image Classification and Retrieval Hybrid Model for Larger Healthcare Datasets using Deep Learning

“…Second, they lack transparency and explainability, which limits the dependability of deep CBIR systems (3,4) . The quality of picture searches by using a strategy that uses multiple deep neural networks and K-Nearest Neighbor algorithms in content-based image retrieval (5) . Recurrent Neural Networks (RNN) are used to automatically generate sentences, whereas Convolutional Neural Networks (CNN) are used to extract the properties of images (6) .…”

“…Because of the number of layers contained in this new approach model, the accuracy % has improved elastically when large datasets are acquired; further the retrieval performance is improved by Euclidean Distance Technique. (5) proposed a method known as Average Precision (AP) and Mean Average Precision (mAP) measures to find the similarity measure of the distance between the feature vectors and conducted model experiments with the Oxford-IIIT Pet Image Dataset and a Kaggle competition self-collected dataset. Accuracy, confusion matrix, and F1 scores are compared with prior models; average precision values are increased with larger deviation.…”