An Efficient Transition Region Based Feature Extraction And Combining Classification Model For Semantic Image Segmentation

Abstract

 SIS (Semantic Image Segmentation) is the process of assigning image pixels to object classes based on previously defined nomenclatures. SIS classifies pixels and is a very demanding task. Existing SIS studies are effective and simple, but they are dependent on robust extractions of required areas. SIS techniques show reduced performances while working with textured backgrounds as gray levels of background and foregrounds overlap. This work proposes improvements in DWTs (Discrete Wavelet Transforms). The scheme called IDWT (Improved DWT) is proposed in this paper to extract features for transition regions and overcomes the issue of background/foreground gray level overlaps. IDWT uses image local variances for getting feature variances where transitional features are separated. Otsu thresholding identifies transitional regions from generated transitional feature images. A morphological operation extracts image edges while filling operations get its object regions. Images are recognized using MLTs (Machine Learning Techniques) while SIS is performed using DLTs (Deep Learning Techniques). EXGB (Enhanced XGBoost) detects super pixels while CNNs (Convolution Neural Networks) trained on monographic images classify objects into SLs (Semantic Labels). The proposed  SIS+ -Q2`method significantly improves semantic segmentation compare to the other image semantic segmentation