الفهرس 
Chapter 1: IntroductionOnly 14 pages are availabe for public view
 |  | from | 245 |  |  |
| | | from | 245 | | |
Abstract
Information retrieval (IR) is the science of searching for information, which has
several forms such as text retrieval, content-based image retrieval (CBIR), and text-to-image
retrieval. Any form of IR, or any feature-based application, mainly depends on the content-
based similarity matching process. For CBIR, it is a standard procedure that measures the
spatial distance between the derived features from a query image and others derived from the
set of training images. It concerns the retrieval of visual information from databases using
image comparison to detect related images in datasets. It is worthy of note that ”image
representation” is the key engine behind the entire process. This is guaranteed across all
proposed test cases for performance evaluation for image matching and retrieval. However,
image representation and consequently similarity matching face several challenges (hurdles),
especially for CBIR. At first, the quantity of detected features and their density must be
compatible with the image content. However, in terms of retrieval accuracy and redundancy,
the number of detected features is scarcely impacted by the quality or sort of the utilised
images. In such cases, redundancy affects not only the size of the applied images but also the
quantum of their extracted visual features. Generated descriptor ”dimensionality” by each
feature extraction method hardly affects matching and retrieval in terms of speed and used
memory. Thus, there is a trade-off between the quality of the applied image and the quantity
of detected features or retrieval accuracy compared to the speed of retrieval. Also, features
extracted are hardly affected by image rotation, shifting, flipping, noise, affine distortion, and
others.
With the massive growth of web images, the scale of databases is enlarged, and new
hurdles are added to the previously mentioned. At first, a suitable deep learning method (CNN
or other) is hardly required to handle large-scale datasets in a general fashion across different
image sorts. In addition, redundancy has an impact across large-scale datasets. Labelling and
indexing are effective, as response time is a key issue in retrieval. For efficient image
representation, especially for sketches, segregate object features as a separate variable across
a large-scale dataset. Also, it is required to avoid recursion because of the local correlation
between pixels. It is highly required to allow model to focus on specific parts of input by
assigning different weights to certain parts of input. Hence, for swift and accurate outcomes,
efficient feature representation of compared images has a major influence on the matching