Contour extraction model introducing contrast adaptive characteristics based on visual pathway

• Published in: Multimedia tools and applications Vol. 84; no. 15; pp. 14693 - 14717
• Main Authors: Fang, Tao, Cai, Zhefei, Fan, Yingle
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.05.2025; Springer Nature B.V
• Subjects: Adaptation; Algorithms; Color imagery; Computer Communication Networks; Computer Science; Contours; Data Structures and Information Theory; Datasets; Image contrast; Image processing; Information flow; Information processing; Multimedia; Multimedia Information Systems; Owls; Retina; Special Purpose and Application-Based Systems; Track 6: Computer Vision for Multimedia Applications; Visual perception; Visual Mechanism; Contrast Adaptation; Dynamic Receptive Field; Contour Detection; Visual Pathway
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-024-19666-y

Effectively extracting image subject contours holds significant importance for subsequent image processing tasks. Recognizing the pivotal role of contrast features in contour characterization, this paper proposes a novel contour extraction model with contrast adaptive characteristics, inspired by visual perception simulation visual pathways. Firstly, leveraging the color antagonism traits of retinal ganglion cells, we construct a color image contrast calculation model to enhance the representation of visual information contrast. Furthermore, drawing upon the feed-forward impact of the retina on Lateral Geniculate Nucleus (LGN) cells, we introduce a contrast gain control mechanism to extract multi-directional primary contour responses. Secondly, employing the LIF neuron network to simulate the lateral geniculate body's information flow processing mechanism, we encode visual information via pulse firing to eliminate redundant contour details. Then, by considering the primary visual cortex (V1) cell's receptive field size variation with contrast, we devise a contrast-adaptive dynamic receptive field model. Additionally, we develop a texture suppression model based on the orientation/contrast disparity between the receptive field model's center and periphery. Finally, integrating the primary contour response and the image obtained through the inhibition model, we fuse features to derive the final image subject target contour. This study employs the BSDS500 dataset and the RuG40 dataset for experimentation. Comparative analysis against mainstream methods demonstrates our approach's superior ability to retain the main image outline while suppressing non-edge textures, with P-value indicators for contour detection of 0.58 and 0.50, respectively. Furthermore, visual analysis using box and whisker plots confirms the robustness of our proposed method.