Research On Target Detection Algorithm Of Marine Benthos Based On Depth Learning

Accurate identification and localization of marine organisms is an indispensable part of marine engineering construction.By monitoring marine organisms,we can better understand the living habits of the target,which is beneficial for the research and protection of marine biodiversity.And it plays an important role in marine aquaculture,marine transportation,and marine autonomous navigation.This article focuses on deep learning-based algorithms for detecting marine biological targets.In response to the problem of small and dense marine benthos,multiscale feature fusion-based small target detection algorithms 3DFPN(Feature Pyramid Network with 3D convolution)are proposed The Transformer based object detection model LSH-DETR(Detection TRansformer with Locally Sensitive Hash attention)and the Loss of Attraction and Repulsion AR-LOSS(Loss of Attraction and Repulsion)based on dense object detection.They verified the effectiveness of marine biological detection on the AUDD dataset and achieved good detection performance in other public datasets(DOTA,URPC).The main research work of this paper is as follows:(1)Research on Object Detection Algorithms and Transformer TechnologyWe conducted in-depth research on classic object detection frameworks and popular visual Transformer models for deep learning-based object detection algorithms.And on this basis,we learned small object detection algorithms based on multi-scale features and object detection losses,compared the advantages and disadvantages of different methods,and provided technical support for deep learning-based marine biological object detection algorithms.(2)Small target detection network of marine benthos based on multi-scale feature fusionIn order to solve the problem of feature extraction of marine benthos with small proportion of pixels,a multi-scale feature fusion small target detection algorithm 3DFPN based on convolutional neural network is proposed.A feature pyramid fusion structure with 3D convolution was designed for small target detection,and redundant feature maps suitable for large and medium targets were removed.Experiments have shown that the improved network structure improves the detection performance of small marine benthos and is suitable for most object detection frameworks.(3)Algorithm for Marine Biological Target Detection Based on TransformerIn order to better utilize the powerful contextual feature information learning ability of advanced popular technology Transformer and improve the accuracy of marine benthos localization and recognition,a new marine benthos detection algorithm LSHDETR is proposed based on DETR.Using a Transformer with locally sensitive hash attention instead of the original Transformer,fully utilizing effective target contextual features,and reducing background information disturbance,in order to achieve more accurate learning of target-related contextual information.In order to be more suitable for target detection in two-dimensional images,relative position encoding has been added on the basis of absolute position encoding.And delete the decoder to accelerate model convergence without affecting detection accuracy.Through comparison and ablation experiments,it has been proven that the proposed LSH-DETR algorithm can achieve a 2.3% improvement in detection accuracy on marine benthos.(4)Target detection loss based on dense marine benthosIn order to improve the dense distribution and mutual occlusion of marine benthos living in groups,the target detection loss was studied and improved,and AR-LOSS was proposed.Firstly,use Focal Loss to address classification errors caused by sample imbalance.Then,drawing inspiration from the idea of gravitational repulsion,a more suitable localization loss for marine dense organisms was proposed: S-Io U loss was used to make the prediction box converge in a direction that is closer to the real box,and two repulsion losses were designed to penalize the prediction box for being close to other real boxes and prediction boxes.The experiment shows that the improved target detection loss improves the detection accuracy of dense marine benthos by 1.7% and reduces the missed detection rate by 1.3%.