The unavailability of reference images in real world problems makes blind image quality assessment (BIQA) a challenging task. The ability of BIQA techniques to assess the image qualityisdirectlydependentonthequalityoffeaturesextracted. ManyBIQAtechniquesare proposed in literature that follow a two-step approach that include extraction of features in different domains and assessment of image quality with the use of extracted BIQA features. TheperformanceofBIQAtechniquescanbedegradedwhenredundantorirrelevantfeatures are present in the image. Therefore, irrelevant and redundant features can be removed using feature selection algorithms that aid in increasing the correlation between predicted quality score and mean observer score (MOS) and lowering the root mean squared error (RMSE), which improves the performance of BIQA techniques. In this thesis, role of feature selection for BIQA has been explored and analyzed. The objectiveoffeatureselectionistoselectfeaturesthatcanhelpinimprovingtheperformance of BIQA techniques. The thesis starts by providing an introduction to image quality assessment followed by a survey of existing state-of-the-art BIQA techniques. The knowledge of existing BIQA techniques is utilized for optimum feature selection, which has not been explored for existing BIQA techniques to the best of our knowledge. In contrast to existing techniques, a three-step framework is presented in this thesis. Existing BIQA techniques are used for feature extraction in the first step. Existing general purpose feature selection algorithms are utilized to reduce the length of feature vector in the second step. The image qualityscoreispredictedutilizingtheselectedfeaturesinthethirdstep. Threeapproachesto feature selection have been considered. Firstly, feature selection is performed using existing feature selection algorithms. During the analysis of features, belonging to various BIQA techniques, it was observed that each distortion type exhibits different characteristics. Each individual distortion type affects each BIQA feature in a distinct manner e.g., Gaussian blur affectsedgeinformationintheimagewhereas,JPEGcompressiondistortiontypeintroduces blockiness in the image. Therefore, using same set of features for all distortion types may not be the optimal approach. Hence, distortion specific feature selection is proposed, which selects different features are selected for each distortion type. Impact of general purpose feature selection algorithms on BIQA techniques has shown promising results. However, thesefeatureselectionalgorithmscanselectirrelevantfeaturesanddiscardrelevantfeatures. Therefore, the performance of fifteen new feature selection algorithms, which are specificallydesignedforBIQA,isexplored. Theproposedfeatureselectionalgorithmsareapplied on the extracted features of existing BIQA techniques and rely on SROCC, LCC, Kendall correlation constant (KCC) and RMSE parameters. Feature selection algorithms based on SROCC and its combination with LCC, KCC and RMSE perform better in comparison to other proposed algorithms. A new BIQA technique based on natural scene statistics properties of the bag-of-features representation and feature selection algorithms is proposed in this thesis. The proposed bag-of-features technique utilizes Harris affine detector and scale invariantfeaturetransformtocomputefeatures, whichareclusteredusingthek-meansclusteringalgorithmtoformthecodebookvocabulary. Thisconstructedcodebookisusedwitha pre-trained support vector regression model to assess the quality of the image. Furthermore, the performance of existing feature selection algorithms is explored on the proposed BIQA technique. Itisobserved,thatfeatureselectionhelpsinimprovingtheperformanceofexistingBIQA techniques,byimprovingtheSROCC,LCC,KCCandRMSEincomparisontousingallthe features for a particular BIQA technique.