Fault detection in seismic profiles is fundamentally constrained by the complex intermixing of geological horizons and noise artifacts, which obscure fault signatures and lead to inaccurate interpretations in traditional detection methods. Conventional coherence-based approaches suffer from limited correlation windows, susceptibility to noise interference, and an inability to preserve fault continuity, particularly under low signal-to-noise ratio conditions. To address these limitations, this study presents an improved ant colony tracking algorithm (IACTA) that integrates connected component labeling with ant colony optimization for enhanced automatic fault detection. By defining connected regions through systematic scanning and labeling, noise regions and fault regions are effectively separated. Seismic profiles are reconstructed as profile-wise 3D attribute surfaces, in which sampling time, trace number, and coherence attribute value jointly define the search space. In this representation, virtual ants perform stepwise shortest-path searching on the attribute surface to trace fault pathways. The proposed profile-wise 3D attribute-surface tracking differs from conventional 2D planar searching because the coherence attribute value is introduced as the third-dimensional constraint for candidate-node selection, heuristic evaluation, and transition-probability calculation. By combining the global search capability of the ant colony optimization algorithm with the spatial perception capability of deep learning models, and utilizing the fault tracking trajectories obtained by IACTA as attention regions for deep learning models, deep learning networks (e.g., U-Net) can be directed to focus on high-probability fault regions, thereby reducing the computational cost of global search and improving feature extraction efficiency. Experimental validation using real-world seismic data demonstrates superior performance compared to established coherence algorithms (C1-C3), with IACTA achieving enhanced fault continuity preservation and effective noise suppression while reducing computational overhead. Parameter analysis reveals minimal sensitivity to most ant colony optimization parameters. The proposed approach offers significant advantages in detection accuracy, noise resistance, and computational efficiency for automated seismic interpretation workflows.