Deposited solder paste inspection plays a critical role in surface mounting processes. When detecting solder pastes defects on a printed circuit board, profile measurement-based methods suffer from large system size, high cost, and low speed for inspection, although they provide 3-D information of solder pastes. In contrast, image analysis-based methods facilitate the defect detection process of solder pastes by treating them as a pattern recognition problem. However, existing image analysis methods do not perform well because low-level visual features cannot catch sufficient information for defect detection. This paper proposes a new defect detection scheme for solder pastes based on learning the color biological feature sub-manifold. In particular, we apply the biologically inspired color feature (BICF) to represent the solder paste images, and introduce a new sub-manifold learning method to extract the intrinsic low-dimensional BICF manifold embedded in an extrinsic high-dimensional ambient space. This scheme mimics the function of human visual cortex in recognition tasks, and can separate poor quality solder pastes from good quality ones. We apply the new scheme to our automated optical inspection system, and thorough empirical studies indicate the effectiveness of the new scheme for practical utilization.