The Non-verbal communication plays a pivotal role in daily life and contributes around 55% to 93 % in overall communications. Facial expression is a type of non-verbal communication and its contribution towards recognition is around 55%. It exhibits the physical intention, behavior, personality and mental state of a person. Facial expression analysis can be effectively used in video surveillance, emotion analysis, smart homes, gesture recognition, patient monitoring, treatment of depression and anxiety, lie detection, automated tutoring, psychiatry, paralinguistic communication, robotics, operator fatigue detection and computer games. Highly accurate solution is a major challenge in the development of Efficient FER system. Data collected using poor quality cameras and/or captured from distance suffer from low resolution problem. Region of interest is usually smaller than original image size and image collected in real world environment suffer from low resolution problem. It results in drastic decrease in classification accuracy of the facial expression recognition. Environmental and source light variations during image acquisition results in poor illumination which is also major cause of performance degradation. Furthermore, curse of dimensionality poses another challenge in the development of fast and accurate techniques. With the increasing demand of surveillance camera-based applications, the Very Low Resolution(VLR) problem happens in many FER application systems. Existing FER recognition algorithms are unable to give satisfactory performance on the VLR face image. In addition to VLR, Variable lighting conditions in uncontrolled environment is another factor which can cause unpredictable illumination affects that leads to poor FER performance. Furthermore, feature vector containing correlated and irrelevant information also causes performance degradation. In this dissertation problems mentioned above related to facial expression recognition have been addressed. A novel framework has been proposed to handle high and low resolution images with equal capability. Excitation component of Weber local descriptor (WLD) is employed to compute the salient features and DWT has been utilized for features extraction which resolves multi-resolution problem. Least number of features having high variances is used to perform classification. Experimental results have shown that this framework not only handle low resolution problem but also gives improved classification performance both in terms of complexity (i.e., number of features) and recognition accuracy as compare to existing techniques present in the literature using CK+, MMI and JAFFE data sets. Secondly, facial expression recognition being a multi-class classification problem is a challenging task and becomes more complicated in real world environment with data having variation in illumination conditions. In order to tackle this problem, illumination invariant technique has been developed based on HOG features. These HOG based illumination invariant features are further reduced using DCT. These highly significant features are passed to the classifiers for accurate facial expression recognition. Proposed frameworks can effectively handle illumination variance and very low resolution data during facial expression recognition. Detail experimentation have been conducted using well known standard datasets containing images with varying illumination, resolution, gender and ethnicity. Comparison of the system has been presented with other state-of-art techniques using CK+, MMI and Cross datasets. Comprehensive experimentation shows that the proposed technique produces significantly better results than existing state-of-the-art techniques present in the related work.