NUST Institutions Library Catalogue

The exponential increase and wide accessibility of visual content on internet have led to
research in the field of image retrieval and search. Image retrieval has been broadly categorized
into content based image retrieval (CBIR) and text based image retrieval (TBIR). TBIR
describes images by manually assigning them words which are processed by a database management
system. However, the manual annotation of images among large databases requires
a lot of time and prone to human error. Whereas, CBIR systems provide automatic retrieval
using texture, shape and color features present in the images. CBIR uses visual content to
identify the similar images.
In this thesis, six CBIR techniques have been proposed to obtain high efficiency and retrieval
precision. Two natural image retrieval techniques have been presented. The first natural
image retrieval technique performs feature extraction based on frequency adder based local
binary pattern (FALBP) and blur detection metric. The extracted features are combined
for accurate image retrieval. The second natural image retrieval technique uses saliency
detection and automatic feature weighting. The technique combines features optimally to
improve the overall time efficiency and precision.
Two techniques for medical image retrieval have been presented. The first technique focuses
on the retrieval of similar retinal disease image. Feature extraction has been done on normal,
choroidal neovascularization (CNV), diabetic retinopathy (DR) and Coats cases. Weights
are assigned to the extracted features using an automatic scheme. The second technique
performs histopathological image retrieval by using multi-channel based decoder concept
and vector of locally aggregated descriptors (VLAD) coding.
Two remote sensing image retrieval techniques have been discussed to address the challenge
of finding interesting content (geographic locations etc) from large number of images. The
first technique achieves building image retrieval by utilizing novel dense angle descriptor
and dictionary learning (DL). The second technique obtains remote sensing based image
retrieval by using dense patch-based local ternary pattern (LTP) and fisher vector coding.
The proposed techniques can be used for biomedical image retrieval to relieve the workload
of doctors and offer a consistent image analysis. Moreover, proposed techniques can also be
used for the retrieval of natural and remote sensing images. Visual and quantitative results
ensure the significance of the proposed techniques. The techniques provide considerably
accurate image retrieval.