Biologically Inspired Hierarchical Contour Detection with Surround Modulation and Neural Connection

Shuai Li, Yuelei Xu, Cong Wei, Shiping Ma, Mingming Zhu, Min QiSensors 2018 · DOI

TL;DR

This paper introduces a novel hierarchical contour detection model inspired by the human visual system. By simulating the complex interplay of neural connections and receptive fields in early visual areas (LGN, V1, V2), the model effectively identifies salient contours in complex scenes while suppressing distracting textures, achieving a good balance between performance and computational efficiency.

What Problem Does This Paper Address?

Detecting contours in natural images is a fundamental yet challenging problem in computer vision. Traditional methods often struggle with complex scenes containing rich textures, clutter, and varying lighting conditions, leading to the detection of spurious edges or the failure to detect important boundaries. While supervised machine learning approaches have shown promise, they can be computationally expensive and heavily reliant on large, well-annotated datasets. This paper aims to bridge this gap by developing a biologically plausible model that leverages the sophisticated processing capabilities of the human visual system to achieve robust and efficient contour detection.

Methods

The proposed Biologically Inspired Hierarchical Contour Detection (BIHCD) model mimics the hierarchical structure of the human visual pathway, from the retina to V1 and V2. Key components include:

  • LGN Layer: Simulates the single-opponent receptive fields of LGN cells to process luminance and color information, creating four distinct channels (red-green and blue-yellow opponent pairs).
  • V1 Layer: Employs simple and complex cell models with orientation-selective receptive fields (modeled using derivatives of Gaussian functions) to extract local edge information from the LGN outputs. A non-classical receptive field (NCRF) mechanism is introduced to achieve surround modulation, which helps suppress internal textures and highlight boundaries. This is achieved by considering luminance, color, and orientation differences between the center and surround regions.
  • V2 Layer: Incorporates feedback connections from V2 to V1, inspired by findings that V2 influences V1's processing of local contours based on global shape information. This feedback mechanism further refines contour detection.
  • Multi-Scale Integration: The model utilizes information from multiple scales to ensure that both fine details and larger structures are captured, and a distance-weighted color-oriented encoding method is used to modulate suppression strength.
  • Neural Connections: The model explicitly simulates feedforward, lateral, and feedback neural connections between visual areas to integrate information and enhance contour detection.

Datasets

The model was evaluated on:

  • Simple Images: Used for initial testing of edge extraction and analysis of V1 CRF characteristics.
  • Complex Human Images: Used to analyze CRF response characteristics and demonstrate NCRF surround inhibitory effects.
  • BSDS300/500 Dataset: A standard benchmark dataset for contour detection, comprising 300 natural images (200 training, 100 test) and an additional 200 test images for BSDS500. This dataset includes multiple human-annotated ground truth contours for each image.

Key Insights and Ideas

  • Hierarchical Processing is Crucial: The paper strongly advocates for a hierarchical approach, mirroring the brain's layered processing, to effectively handle complex visual scenes.
  • Surround Modulation is Key for Texture Suppression: The non-classical receptive field (NCRF) mechanism, which considers contextual information beyond the classical receptive field, is vital for distinguishing salient contours from distracting textures.
  • Feedback Connections Enhance Contour Salience: The inclusion of V2 feedback to V1 demonstrates how global shape information can refine and reinforce local contour detection.
  • Multi-Feature Integration Improves Robustness: Combining luminance, color, and orientation information, along with multi-scale processing, leads to more robust contour detection, especially in challenging conditions.
  • Biological Plausibility Offers Performance Gains: By drawing inspiration from neurophysiological findings, the model achieves competitive or superior performance compared to existing methods, often with lower computational cost.

Key Results

  • Improved Contour Detection: The BIHCD model achieved a good trade-off between edge detection and texture suppression, outperforming several state-of-the-art biologically-inspired and machine learning methods on the BSDS300/500 dataset.
  • Quantitative Performance: The paper reports F-measure and Average Precision (AP) values, showing significant improvements over methods like Canny and Pb, and competitive results against deep learning models. For instance, on BSDS500, the BIHCD model achieved an F-measure of 0.70 and AP of 0.74, compared to Human performance of 0.80 and 0.81 respectively.
  • Computational Efficiency: The model is significantly faster than some deep learning algorithms, with a mean computation time of 15 seconds on MATLAB R2017a, compared to 133s for gPb and 213s for HED.
  • Parameter Robustness: The BIHCD model requires optimization of only two main parameters, indicating higher robustness compared to methods with numerous unknown parameters.

Limitations and Future Work

  • Low-Level Focus: The current model primarily focuses on low-level, local visual information. The authors acknowledge that visual cells exhibit stronger dynamic perception characteristics and that higher-level cells (like V4) have special feedback effects on low-level cells.
  • Future Directions: Future work will explore rapid contour extraction driven by target information and global features in complex scenes. The authors also plan to integrate a higher level of visual features to improve contour detection performance and extend its application to other computer vision tasks.

Summary updated

Abstract

Contour is a very important feature in biological visual cognition and has been extensively investigated as a fundamental vision problem. In connection with the limitations of conventional models in detecting image contours in complex scenes, a hierarchical image contour extraction method is proposed based on the biological vision mechanism that draws on the perceptual characteristics of the early vision for features such as edges, shapes, and colours. By simulating the information processing mechanisms of the cells' receptive fields in the early stages of the biological visual system, we put forward a computational model that combines feedforward, lateral, and feedback neural connections to decode and obtain the image contours. Our model simulations and their results show that the established hierarchical contour detection model can adequately fit the characteristics of the biological experiment, quickly and effectively detect the salient contours in complex scenes, and better suppress the unwanted textures.
BibTeX
@inproceedings{li2018biologically,
  author = {Shuai Li and Yuelei Xu and Cong Wei and Shiping Ma and Mingming Zhu and Min Qi},
  title = {{Biologically Inspired Hierarchical Contour Detection with Surround Modulation and Neural Connection}},
  booktitle = {Sensors},
  year = {2018},
  doi = {10.3390/s18082559},
}
Figure