基于声呐的水下目标检测是具有重要意义的研究课题。由于声呐图像质量差、对比度低、边缘模糊等，基于特征提取的识别方法在精度和速度上无法满足现有要求。鉴于此，将基于卷积神经网络的 YOLOv3 引入声呐图像的目标识别任务中，首先通过动态亮度分配和中值滤波对图像进行预处理建立数据集，然后对 YOLOv3 模型进行训练和测试，最后根据静态目标的位置不变去除虚假目标，降低误检率。实验结果表明： 较之现有的特征提取识别算法，本算法具有更高的效率和更低的误检率。
Underwater target detection based on sonars is an important research subject. Due to the low quality，low contrast and blurred edges of sonar images，the recognition method based on feature extraction cannot meet the current requirements in terms of accuracy and speed. In view of this，this paper introduces YOLOv3 based on convolutional neural network(CNN)into the target recognition task of sonar images. First，the images are preprocessed by dynamic brightness distribution and median filtering to establish a data set，then the YOLOv3 model is trained and tested，and finally the false detection rate is reduced through eliminating false targets according to the unchanged location of the static target. The test results showed that compared with the existing feature recognition algorithm，this algorithm has higher efficiency and lower false detection rate.