[Paper Review] Detection of curved lines with B-COSFIRE filters: A case study on crack delineation

Nicola Strisciuglio, George Azzopardi+1 more|arXiv (Cornell University)|Jul 24, 2017

Retinal Imaging and Analysis19 references4 citations

TL;DR

This paper proposes a brain-inspired B-COSFIRE filter pipeline for detecting curved, thin, and incomplete lines in noisy images, achieving state-of-the-art performance (F-measure = 0.865) on the Crack_PV14 dataset. The method combines nonlinear filtering, non-maximum suppression, hysteresis thresholding, and morphological closing to robustly delineate curvilinear structures such as cracks in pavement images.

ABSTRACT

The detection of curvilinear structures is an important step for various computer vision applications, ranging from medical image analysis for segmentation of blood vessels, to remote sensing for the identification of roads and rivers, and to biometrics and robotics, among others. %The visual system of the brain has remarkable abilities to detect curvilinear structures in noisy images. This is a nontrivial task especially for the detection of thin or incomplete curvilinear structures surrounded with noise. We propose a general purpose curvilinear structure detector that uses the brain-inspired trainable B-COSFIRE filters. It consists of four main steps, namely nonlinear filtering with B-COSFIRE, thinning with non-maximum suppression, hysteresis thresholding and morphological closing. We demonstrate its effectiveness on a data set of noisy images with cracked pavements, where we achieve state-of-the-art results (F-measure=0.865). The proposed method can be employed in any computer vision methodology that requires the delineation of curvilinear and elongated structures.

Motivation & Objective

• To develop a general-purpose, robust detector for curvilinear structures in noisy and complex images.
• To address the challenge of detecting thin, incomplete, and tortuous lines in real-world images, such as cracks in pavement.
• To leverage brain-inspired, trainable filters that require no manual feature engineering and generalize across varying line shapes and orientations.
• To achieve state-of-the-art performance on benchmark crack detection datasets without requiring extensive ground truth for training.
• To provide a flexible, modular pipeline applicable to diverse computer vision tasks involving line segmentation.

Proposed method

• Apply B-COSFIRE filters—trainable, brain-inspired filters modeled after simple cells in V1 cortex—that are configured automatically from a prototype pattern.
• Use nonlinear filtering with B-COSFIRE to respond selectively to elongated, curved structures of specific widths and orientations.
• Apply non-maximum suppression to thin the response map into single-pixel-wide lines.
• Apply hysteresis thresholding to preserve weak responses connected to strong ones, enhancing continuity.
• Perform morphological closing to reconnect broken line segments and fill small gaps.
• Control robustness via parameters σ₀ and α, enabling tolerance to variations in line shape and noise.

Experimental results

Research questions

• RQ1 Can a trainable, brain-inspired filter approach outperform existing methods in detecting thin, incomplete, and tortuous curved lines in noisy images?
• RQ2 How robust is the B-COSFIRE-based pipeline to noise, line discontinuities, and high tortuosity in real-world images?
• RQ3 To what extent can the B-COSFIRE filter’s automatic configuration and representation learning reduce the need for manual feature engineering?
• RQ4 Does the proposed pipeline achieve state-of-the-art performance on benchmark crack detection datasets like Crack_PV14?
• RQ5 Can the method be generalized across diverse computer vision applications requiring curvilinear structure detection?

Key findings

• The proposed method achieves a state-of-the-art F-measure of 0.865 on the Crack_PV14 dataset, outperforming existing approaches.
• The B-COSFIRE filter demonstrates high robustness to noise, with signal-to-noise ratios (SNR) as low as 3.56 dB in tested stimuli.
• The method successfully detects incomplete lines and highly tortuous structures, such as those following a 1D Gabor function, even under significant noise.
• The pipeline maintains high accuracy across varying line widths and orientations due to the filter’s trainable and tolerant design.
• The average processing time per image (200×300 pixels) is 0.3575 seconds on a 2.7 GHz processor, with potential for parallelization.
• The method is robust to line discontinuities, as shown by consistent detection of dashed circular patterns with angular gaps up to 5°.