CNN-EFF: CNN Based Edge Feature Fusion in Semantic Image Labelling and Parsing

摘要

Semantic segmentation and image parsing have rapidly become an eminent research area in computer vision and machine learning domain. Many applications have required a robust mechanism for segmentation, such as self-driving, augmentative reality, object recognition, etc. Due to the high applicability in the various domains, In this paper, we have introduced a two-step frame-work that parses the image into predefined labels by using a novel CNN architecture and improving the likelihood of labels. In step-1, nine-layer CNN architecture has been introduced, which trains on minimal training samples and results in the pixel-wise Soft-Max probabilities. These probabilities are the soft estimates derived from a hard classifier, i.e., MLP. Data in step-1 has been prepared in the form of a patch-label set. In step-2, we have introduced a Jacobian optimization-based label relaxation method that fuses the local extrema as an edge prior. The proposed frame-work has been denoted as CNN-EFF in this work. The CNN-EFF scheme has been evaluated two publicly available benchmark data-sets, which has arranged in the form of image and their pixel label ground-truth. The experimental results have been compared with the previously proposed state-of-the-art methods. The CNN-EFF has greatly improved semantic labeling accuracy up to a significant gain from the past techniques. The CNN-EFF process has reported 84.42%, 85.91%, 94.66%, 97.14%, and 98.27% accuracy for the Highway, House, sheep, Horse-rider, and Horse-keeper images, respectively. Conclusively, the Proposed frame-work has out-performed the previously proposed state-of-the-art methods.