基于改进Faster R-CNN的水母检测与识别算法

提出一种基于改进Faster R-CNN水母检测与识别算法。首先,建立了包含7种水母的数据集;然后,针对ResNeXt(C=32)用于目标检测时出现计算量较大的问题,在保证精确度的前提下,将分支数C设置为8以降低计算量;最后,为解决水母检测时出现的检测精度低和小个体无法检测的问题,在残差网络中引入膨胀卷积。实验结果表明:该算法较VGG16、ResNet101、ResNeXt(C=32)和ResNeXt(C=8)方法,mAP值分别提高了3.15%、2.09%、3.01%和2.36%;F1-score分别提高了2.53%、1.99%、2.01%和2.31%;loss损失函数收敛值更优,收敛精度趋近于0。P-R曲线、可视化效果分析和水母视频检测的结果证明:该算法的水母检测准确率和水母检测数量明显优于其他算法,检测精度较高,基本可以达到实时监测的要求。

Jellyfish Detection and Recognition Algorithm Based on Improved Faster R-CNN

A jellyfish detection algorithm based on improved Faster R-CNN is proposed. Firstly, a data set containing 7 species of jellyfishes is established. Secondly, on the premise of ensuring the accuracy, the number of branches C is set to 8 to solve the problem that ResNeXt (C=32) has a high amount of calculation for target detection. Finally, to solve the problems of low detection accuracy and small individuals unable to be recognized, expansion convolution is introduced into the residual network. The experimental results shown that compared with VGG16, ResNet101, ResNeXt (C=32) and ResNeXt (C=8), the mAP value of the proposed algorithm increase by 3.15%, 2.09%, 3.01% and 2.36%. F1-score increase by 2.53%, 1.99%, 2.01% and 2.31%. Loss function convergence value of the proposed algorithm approach to 0. Results of P-R curve, visual analysis and video detection show that the accuracy and detection number of jellyfish by the proposed algorithm is the best, the proposed algorithm has high detection accuracy and can meet the requirements of real-time monitoring.