Hybrid deep learning and active contour for segmenting hazy images

Abstract

Image segmentation seeks to distinguish the foreground from the background for further analysis. A recent study presented a new active contour model (ACM) for image segmentation, termed Gaussian regularization selective segmentation (GRSS). This interactive ACM is effective for segmenting certain objects in images. However, a weakness of the GRSS model becomes apparent when utilized on hazy images, as it is not intended for such conditions and produces inadequate outcomes. This paper introduces a new ACM for segmenting hazy images that hybridizes a pretrained deep learning model, namely DehazeNet, with the GRSS model. Specifically, the haze-free images are estimated using DehazeNet, which fuses the information with the GRSS model. The new formulation, designated as GRSS with DehazeNet (GDN), is addressed via the calculus of variations and executed in MATLAB software. The segmentation accuracy was evaluated by calculating Error, Jaccard, and Dice metrics, while efficiency was determined by measuring processing time. Despite the increased processing time, numerical experiments demonstrated that the GDN model achieved higher accuracy, as indicated by the lower error and higher Jaccard and Dice than the GRSS model. The GDN model can potentially be formulated in the vector-valued image domain in the future.