Aiming at the problem that Faster R-CNN is prone to miss detection of small targets in multi-scale target detection, an improved multi-scale target detection algorithm is proposed. Firstly, the low-level features used for small target detection are fused with high-level features used for large target detection. Secondly, an online hard example mining algorithm is applied to maintain a hard case sample classification pool for accelerating the iterative convergence of the neural network model, which could solve the problems includes uneven training samples and low training efficiency. Finally, the size of the detected target is calculated and counted for controlling the size of the anchor box and improving the generalization ability of the model. The experimental results show that compared with Faster R-CNN, the mean average precision (mAP) is improved by 8.61 and 5.47 percentage points respectively.