Printed circuit board plays a pivotal role in the development of industry. Automatic detection of targets in printed circuit board can greatly save the human resources and enhance the speed of detection. However, due to the inherent characteristic of printed circuit board, the targets are usually small objects with various illumination and shapes. We propose a general architecture, named ACTDet, to tackle the problems of small objects detection in printed circuit board. In specific, we design a cross grained transformer to leverage the low-level and high-level features and enhance the inherent interaction of different feature maps with various receptive fields. A number of data augmentation strategies are adopted to enrich training data, avoid overfitting, and improve robustness of detector. What’s more, a channel-guided attention block is utilized to reweight the contribution of each channel and encourages the network to learn the most important features. Extensive experiments illustrate the effectiveness of the proposed ACTDet network and ACTDet achieves state-of-the-art performance in our proposed well-annotated dataset.