In order to realize the accurate quantitative detection of marine benthos and solve the problems in detecting small and densely distributed benthic organisms under overlapping and occlusion image, a quantitative detection algorithm for marine benthos based on Faster R-CNN is proposed. A convolution kernel adaptive selection unit is embedded in the backbone to enhance the feature extraction ability of network. Based on this, multi-resolution feature fusion is introduced to design deconvolution feature pyramid structure for small object detection. At the same time, the selection of anchor in Region Proposal Network is optimized to improve the accuracy of counting. Transfer learning strategy is also employed to train the proposed model and alleviate the limitation of small dataset. The results show that compared with the original Faster R-CNN, the proposed algorithm improves the recognition precision of marine benthos from 93.25% to 96.32%, and reduces the mean average error from 16.53 to 7.38. This improvement reflects that the proposed algorithm is more suitable for the quantitative detection of small and dense objects on the seafloor.

为了实现对海洋底栖动物的准确定量检测，解决重叠和遮挡图像下小而密集分布的底栖生物的检测问题，提出了一种基于Faster R-CNN的海洋底栖动物定量检测算法。卷积核自适应选择单元嵌入主干中，以增强网络的特征提取能力。在此基础上，引入多分辨率特征融合，设计反卷积特征金字塔结构，用于小目标检测。同时，对区域投标网络中锚点的选择进行了优化，以提高计数的准确性。转移学习策略也被用来训练提出的模型并减轻小数据集的局限性。结果表明，与原始的Faster R-CNN相比，提出的算法将海洋底栖动物的识别精度从93.25％提高到96.32％，平均平均误差从16.53降低到7.38。这种改进反映出，所提出的算法更适合定量检测海底小而密集的物体。