Image retrieval with SNN-based multi-level thresholding

Image retrieval is defined as indexing similar or identical images in a digital image database. Various feature vectors obtained from the images are used while searching for a similar digital image. However, processing all pixels of the images requires costly algorithms. In addition, it is a possible issue that the images used in retrieval approaches are of different sizes. For this reason, pixel-level operations are insufficient when comparing images. Therefore, it requires vectorial structures that represent images. The process of obtaining these vectorial structures is called feature extraction, and it is one of the most important stages of content-based image retrieval. On the other hand, the histogram is the most basic feature vector that is independent of the dimensions of the image and can be easily calculated. In gray-level images, the size of the histogram is suitable for use as a feature vector. However, three different channels in color images contain too much data to be used as feature vectors. The data of 3 separate histograms are reduced using various thresholding processes and feature vectors are extracted. Therefore, reducing the vector size is an inevitable operation. In this study, a new multi-thresholding method based on the Spiking Neural Network model, inspired by the human visual system, is proposed. With the proposed model, 3 threshold values are determined for each of the RGB color channels, and each color channel is divided into 4 parts. Thus, the color palette of the image is quantized to 64 different colors and a feature vector with 64 elements is obtained. The proposed method was compared with the commonly used multilevel thresholding