改进的Faster R-CNN海洋鱼类检测模型

Improved Faster R-CNN Marine Fish Detection Model

为提高海洋鱼类检测的准确率,提出一种基于Faster R-CNN的海洋鱼类检测方法.首先,利用迁移学习方法训练Faster R-CNN网络,克服海洋鱼类样本集有限的限制;其次,增加颈部连接层,使用双向特征金字塔网络(BiFPN)进行特征融合,得到具有丰富位置信息和语义信息的融合特征图;再次,将卷积层输出的特征矩阵作外积相乘运算,提高对相似海洋鱼类的识别精度;最后,结合Mask R-CNN中的ROI Align方法对预测位置进行二次修正,使检测框更加准确.实验结果表明,在检测海洋鱼类数据集时,与原始的Faster R-CNN算法相比,改进后的Faster R-CNN检测模型的平均准确度均值提高了7.4%.

In order to improve the accuracy of marine fish detection,a Faster R-CNN based detection method for marine fish is proposed.Firstly,the Faster R-CNN network is trained by using the migration learning method to overcome the limited sample set of marine fishes.Secondly,the neck connection layer is added and the bi-directional feature pyramid network(BiFPN)is used to fuse the features.The fusion feature map with rich position information and semantic information is obtained.Then the feature matrix output from the convolution layer is multiplied by the outer product to improve the recognition accuracy of similar marine fishes.Finally,the ROI Align method in Mask R-CNN is combined.In order to make the detection frame more accurate,align method modifies the prediction position twice.The experimental results show that compared with the original Faster R-CNN algorithm,the average accuracy of the improved Faster R-CNN detection model is improved by 7.4%in the marine fish data set.