In recent years, object detection algorithms have achieved great success in the field of machine vision. To pursue the detection accuracy of the model, the scale of the network is constantly increasing, which leads to the continuous increase in computational cost and a large requirement for memory. The larger network scale allows their execution to take a longer time, facing the balance between the detection accuracy and the speed of execution. Therefore, the developed algorithm is not suitable for real-time applications. To improve the detection performance of small targets, we propose a new method, the real-time object detection algorithm based on transfer learning. Based on the baseline Yolov3 model, pruning is done to reduce the scale of the model, and then migration learning is used to ensure the detection accuracy of the model. The object detection method using transfer learning achieves a good balance between detection accuracy and inference speed and is more conducive to the real-time processing of images. Through the evaluation of the dataset voc2007 + 2012, the experimental results show that the parameters of the Yolov3-Pruning(transfer): model are reduced by 3X compared with the baseline Yolov3 model, and the detection accuracy is improved, realizes real-time processing, and improves the detection accuracy.

近年来，目标检测算法在机器视觉领域取得了巨大的成功。为了追求模型的检测精度，网络规模不断扩大，导致计算成本不断增加，对内存的需求也很大。更大的网络规模使得它们的执行时间更长，面临检测精度和执行速度之间的平衡。因此，所开发的算法不适合实时应用。为了提高小目标的检测性能，我们提出了一种新的方法，即基于迁移学习的实时目标检测算法。在基线 Yolov3 模型的基础上，进行剪枝以减小模型的规模，然后使用迁移学习来保证模型的检测精度。使用迁移学习的目标检测方法在检测精度和推理速度之间取得了很好的平衡，更有利于图像的实时处理。通过对数据集 voc2007 + 2012 的评估，实验结果表明，Yolov3-Pruning(transfer): 模型的参数与基线 Yolov3 模型相比减少了 3 倍，提高了检测精度，实现了实时处理, 提高了检测精度。