Automated printed circuit board signal testing is widely used for quality testing and functional analysis. Segmentation-based visual technology provides the core localization and recognition capability for automated signal testing systems. CNN-based segmentation models achieve significant improvement over the past few decades. However, there are still two key difficulties in the PCB test solder joints segmentation task, i.e., dense distribution and extremely small size, which make it hard for CNN models to obtain high-accuracy localization and recognition of small objects. In this paper, we build a printed circuit board segmentation dataset (PCBSeg) for training the PCB segmentation models. We propose a novel dense feature augmentation network, named DFANet, to strengthen the feature representation ability of small target objects. We exploit the attention and transformer that benefit modeling long-range feature relationship to fuse multi-scale features and enhance the information of dense small target features. Extensive experiments illustrate that our proposed DFANet achieves the state-of-the-art performance on the PCBSeg dataset.