Search

Segmenting precisely affected parts of the lungs from the output of CT (Computed Tomography) is critical in making inquiries on lung malignancy and can offer significant data for clinical conclusions. It plays a major and effective role in researches on lung diseases. The crux of the problem is developing automatic detection of lesion and segments them with perfect accuracy. Heterogeneity of lesion part makes segmentation a very difficult task. In TBGA (Toboggan Based Growing Automatic Segmentation Approach), the lack of degree of recognition results in difficulty in the boundary detection process during segmentation. To overcome the drawbacks, a Regression Neural Networks (RNN) Segmentation approach has been proposed in this paper. The degree of recognition is less in tissues which are associated to the neighboring lesion with pixels having same intensity. RNN provides a greater accuracy of recognition of the adjacent lesions with similar intensity when compared to other methods like Skeleton graph cut and Level set method. Segmentation is done based on the degree of recognition. So the RNN method proposed in this paper concentrates mainly on the precise detection of boundary for juxtapleural and juxtavascular lesions. The accuracy of segmenting lung parenchyma is a challenge in lesion segmentation. In RNN segmentation process, the result of parenchyma forms the basis of extracting lesion. RNN is a Learning Algorithm so the complexity of automatic lesion detection is avoided. RNN uses a trained set of data, so the resulting outcome is accurate.

Local descriptors play a vital role in face analysis. This paper proposes a modified dimensionality reduced local directional pattern (MDR-LDP) for face analysis that includes both face and facial expression recognition. MDR-LDP is an updated version of dimensionality reduced local directional pattern (DR-LDP) for face recognition. DR-LDP assigns a single code for every $$3 \times 3$$ region of local directional pattern (LDP) encoded image. It is a compact and efficient code to recognize faces but gives only satisfactory results for facial expression recognition. Every $$3 \times 3$$ LDP encoded region is reduced to a $$2 \times 2$$ region and labeled as MDR-LDP. The MDR-LDP descriptor considers the micro structure patterns in the image. The resultant MDR-LDP encoded image is further divided into regions and histograms are generated for each region. The bins of each histogram form the feature vector. The concatenation of all the feature vectors forms the MDR-LDP descriptor and is used for facial expression. The proposed MDR-LDP is robust to noise, illumination changes and pose variations. The experiments were carried out on standard databases and the objective evaluation using the standard metrics prove that MDR-LDP performs superior to the other local descriptors for face analysis.