Nowadays, high-frequency forward-looking sonar is an effective device to obtain the main information of underwater objects. Detection and segmentation of underwater objects are also one of the key topics of current research. Deep learning has shown excellent performance in image features extracting and has been extensively used in image object detection and instance segmentation. With the network depth increasing, training accuracy gets saturated and training parameters also increase rapidly. In this paper, a series of residual blocks are used to build a 32-layer feature extraction network and take place of the Resnet50/101 in Mask RCNN, which reduces the training parameters of the network while guaranteeing the detection performance. The parameters of the proposed network are 29% less than Resnet50 and 50.2% less than Resnet101, which is of great significance for future hardware implementation. In addition, Adagrad optimizer is introduced into this research to improve the detection performance of sonar images. Finally, the object detection results of 500 test sonar images show that the mAP is 96.97% that is only 0.18% less than Resnet50 (97.15%) but more than Resnet101 (95.15%).

如今，高频前视声纳是一种获取水下物体主要信息的有效装置。水下物体的检测和分割也是当前研究的关键主题之一。深度学习在图像特征提取中表现出出色的性能，并已广泛用于图像对象检测和实例分割。随着网络深度的增加，训练精度变得饱和，训练参数也迅速增加。本文使用一系列残差块构建了一个32层的特征提取网络，并取代了Mask RCNN中的Resnet50 / 101，在保证检测性能的同时，减少了网络的训练参数。拟议网络的参数比Resnet50小29％，比Resnet101小50.2％，这对于将来的硬件实现具有重要意义。此外，本研究还引入了Adagrad优化器以提高声纳图像的检测性能。最后，500个声纳图像的目标检测结果表明，mAP为96.97％，仅比Resnet50（97.15％）小0.18％，但比Resnet101（95.15％）大。