Article Preview
TopIntroduction
With the advent of the information technology era, the use of image retrieval in various fields, such as medicine, entertainment, and multimedia, has increased. The required number and size of images are increasing rapidly. Thus, the speed and accuracy improvement of image retrieval has become a popular research topic. Among the existing image retrieval technologies, content-based image retrieval (CBIR) (Feng, L et al., 2015; Li, W. et al., 2017; Pan, H. et al., 2013) has been preferred by researchers. Unlike text-based image retrieval technologies (Xin, G, U et al., 2014), CBIR does not use the text description of an image to conduct a search but carries out a sample image retrieval from the image library according to the low-level features of the image (Khatabi, A. et al., 2014). However, the results of image retrievals using low-level features are unsatisfactory. Two methods have been proposed to solve this problem. One is the relevance feedback technique, which was introduced into image retrieval to improve the quality of retrieval (Veganzones, M. A et al., 2013). However, users are required to provide feedback on the retrieval results, thereby resulting in multi-iteration that increases the time overhead of retrieval. In the other method, the low-level features of images are accorded before image retrieval for the effective classification of images in the library using cluster algorithm (An, Y. et al.,2016). Thus, the retrieval scope can be reduced to improve retrieval accuracy and speed.
Clustering algorithms are used to classify similar data (Nayak, P. et al., 2015) and are widely applied in image retrieval and other fields. The Fuzzy C-mean algorithm generates the initial cluster center using a random number (Wu, X. et al., 2015). The number of generated clusters and the misclassification rate are high. These conditions are suitable for data clustering with spherical structures but not for clustering image data. K-means is the most commonly used clustering algorithm in applications. However, its clustering results depend on the selection of initial cluster centers, and it is sensitive to noise and abnormal data in data set, which results in difficulty in generating stable clustering results (Kurt Hornik et al., 2017). Thus, when K-means algorithm is used to cluster image data, the clustering results exhibit strong randomness. PAM clustering algorithm is an improved algorithm based on the K-means clustering algorithm. The PAM algorithm selects data objects as cluster centers (Li, Z. et al., 2017). Thus, the clustering result of the PAM algorithm is better than that of the K-means algorithm when noise and abnormal data exist in the data set. However, the amount of computation time required by the PAM algorithm is larger than that in finding cluster centers, thereby severely affecting the clustering time.
An improved PAM algorithm is proposed based on the advantages and disadvantages of the above algorithms. The improved particle swarm algorithm-fireworks particle swarm algorithm (PSO-FWA) can optimize PAM algorithm’s process of identifying cluster centers and improve its clustering efficiency. The algorithm is used in clustering feature vectors and exhibits improved clustering results and image retrieval accuracy. The simulation experiment was carried out by applying the improved PAM algorithm and K-means clustering algorithm proposed in this paper to image retrieval. It can be seen that the image retrieval method proposed in this paper is more accurate.
We make the following contributions in this work:
- 1.
Optimize the particle swarm algorithm through the firework algorithm, and propose a new optimized particle swarm algorithm, namely the firework particle swarm algorithm (PSO-FWA algorithm).
- 2.
The PAM algorithm is optimized by the PSO-FWA algorithm, and a new clustering algorithm is proposed, namely the PF-PAM algorithm, to improve the clustering effect.
- 3.
Preprocess the image database data through the PF-PAM algorithm to improve the accuracy of image retrieval.