<p>Current approaches in pixel-level image analysis often rely on uniformly distributed receptive fields, neglecting the critical role of pixel-specific feature density, which may lead to insufficient local feature representation. To address this, we introduce the Local Dilated Convolution (LDC) module and its lightweight variant, Ghost-LDC. LDC combines a standard local convolution branch with a dilated convolution branch within the same operator. This intra-layer compensation improves the continuity of the theoretical receptive-field support and, as observed in the support visualization, produces a center-enhanced Gaussian-like response tendency, rather than relying only on layer-wise dilation-rate scheduling. Ghost-LDC further incorporates the Ghost module to reduce the additional computational burden introduced by the full LDC branch. Experiments on floc feature detection, crowd counting, and semantic segmentation show that LDC-based modules can improve performance in some tested dense prediction settings, while the semantic segmentation results are more configuration-sensitive. Ghost-LDC provides a more practical trade-off between accuracy and computational cost.</p>