Research On Underwater Biometrics Based On Deep Learning

Underwater biometrics technology is to accurately identify underwater creatures in images and videos through computer algorithms.Based on YOLOv5 and Cascade R-CNN as the basic underwater target recognition models,this paper studies the underwater biometrics technology based on deep learning,and designs the underwater biometrics model Fattention-YOLOV5 and the small-scale underwater biometrics model AIMC R-CNN to improve the recognition accuracy.Based on the above technology,an underwater biological detection application system is realized.The specific research is as follows:First,the underwater biometrics model Fattention-Yolo V5 is designed.Based on YOLOv5 model,an F-CBAM attention mechanism was proposed.By referring to FRe LU activation function in CBAM structure,complex feature distribution was captured in 2d space at the activation function stage to achieve pixel-level spatial information modeling ability and improve model accuracy.The channel attention mechanism and spatial attention mechanism in F-CBAM are used to improve the channel weight of the target object and expand the receptive field of the target to the original image to improve the feature learning ability of the target detection model.Secondly,bidirectional cross-scale feature pyramid Bi FPN was added to improve the identification accuracy of underwater organisms through bidirectional cross-scale connection and weighted image feature fusion.Finally,KL LOSS function was used to reduce the influence of ambiguous boundary boxes on the prediction results and improve the recognition rate of the concealed underwater organisms.Experimental results show that the designed Fattention-Yolo V5 has a high accuracy.Second,a small scale underwater biometrics model AIMC R-CNN is designed.Based on the Cascade R-CNN model,the image enhancement method of exposure balance is used to improve the color contrast and detail information of underwater images,which lays a foundation for improving the recognition rate.Res Ne Xt was added as the image feature extraction network to obtain the underwater biological target feature map with higher accuracy.Finally,the multi-scale feature aggregation interaction strategy is adopted to improve the expression ability of image features at different network layers and improve the recognition accuracy of the model for small targets.The experimental results of AIMC R-CNN were analyzed,the results of images enhanced by the exposure balance image enhancement method were compared,and the ablation experiments of different improved points were carried out.Finally,compared with the mainstream recognition model,the designed AIMC R-CNN model was three percentage points higher than the original Cascade R-CNN m AP@0.5.The effectiveness of the improved method is proved.Finally,the underwater biometrics model AIMC R-CNN based on deep learning is integrated into the underwater biometrics system.The test results of the underwater biometrics system show that the designed underwater biometrics detection system has high accuracy and usability,and can be used to identify biological targets in the underwater real-time environment,providing a more solid and accurate data foundation for Marine scientific research.