Object detection performs a vital function in numerous domains. In contemporary object detection frameworks, multiple detection heads are a crucial element. Multi-performance head improvement is mostly attributable to the employment of many heads to detect objects of different sizes, but this also results in significant computing usage. Moreover, not all heads need to participate in the detecting process for every image. We first investigate the correlation between the number of heads and computing consumption, explore how to determine whether a detection head is in charge of detecting an object, and propose a dynamic head network for detection (DH-Net), which according to the input photographs, adaptively pick which heads to make predictions. As a result, the technique can accomplish more effective dynamic reasoning and a better balance between detection accuracy and computing expense. DH-Net allows the model to lower the number of parameters by up to 34% while still maintaining comparable accuracy, according to a series of controlled trials in the TT100K datasets. Moreover, it has the effect of speeding up training for redundant models.