<![CDATA[The research aims to determine the nature and features of a digital image in order to suggest or recommend relative images (candidate) to the image (query) that is determined by the user which may contain high scientific or valuable contents. The proposed method processes the semantic segmentation firstly, then computing the feature similarity, by segmenting the image into meaningful object regions then extracting isolating features for the determined objects, then the system will enhance the matching range between the recommendation query and the output of the semantic level. The sample was consisting of 1380 images based on 70/30 split with selected 3 labels (cars, persons and trees). The preprocessing step has been applied where the features were extracted from the determined image in order to determine its contents then we can recommend (candidate) the images with the relative features, this process is applied using the concept of semantic segmentation, where the procedure is partitioned into several steps: The first step includes grouping (cars, persons and trees) images from multiple online and real-world datasets, the grouped raw data that can be processed by the system normalization have been processed using Min-Max normalization that is used for standardizing the input data. The second step is the feature extraction process which is achieved using a modified VGG-16 net and the fully connected layers have been removed and the output will be a 2-D features map, then we convert this map to a 1-D vector called feature vector using a Global Average Pooling. The next step is obtaining feature labels by applying a Support Vector Machine SVM classifier, then we recommended the images with relative features to the determined (query) image, this step is achieved using a Pearson correlation coefficient. The final step is constructing a Confusion Matrix and applying the (F1-score, Recall, Precision and Accuracy) metrices in order to estimate the performance.]]>
Copyrights © 2026
