Research On Wildlife Target Detection Based On Deep Learning

Wild animals are one of the important components of natural ecosystems,and it is of far-reaching significance to realize the rapid and accurate identification of wild animals in nature reserves through deep learning algorithms.Aiming at the problems of low recognition rate and time-consuming manual labeling of camera traps,this dissertation studies the wild animal detection model,and realizes the accurate detection of wild animals in high-resolution and complex scenes.The main research contents of this dissertation are as follows:(1)A wild animal detection algorithm based on YOLOv5-CA is proposed.The YOLOv5 algorithm is improved.Based on the YOLOv5 s model,a multi-scale feature detection layer is added,and the coordinated attention mechanism is introduced.By aggregating vertical and horizontal position sensing features,the network structure can obtain more receptive fields.The experimental results show that the accuracy of the original YOLOv5 s is 91.2%,and the accuracy of the improved YOLOv5-CA is 96.8%,which is increased by 5.6%.The detection of each classification is significantly improved.YOLOv5-CA can be more accurate and better improve the performance of wild animal detection.(2)The wild animal target box annotation software Auto Label YOLOv5-CA was designed.Aiming at the problem that the Enonkishu dataset has not yet published data annotation files and the dataset still uses human annotation methods.By constructing multiple single-classified wildlife monitors,analyzing the YOLOv5-CA algorithm network model,and batch labeling the dataset.It solves the problem of missing,error-prone,and inaccurate anchor boxes in the human-labeled dataset,while reducing the repetitive workload.(3)Research on the Swin Transformer network model and propose the HESWTR network.A Hybrid Enhanced Block is proposed in the backbone network,which is embedded between Patch Merging and Swin Transformer Block to enhance the information interaction between the shifted window and the context,enhance the receptive field,and improve the feature extraction capability of the model.The high-resolution feature pyramid network(HRFPN)is introduced into the NECK layer of the network model to enhance the feature extraction capability of wild animals at high-resolution scales.In order to solve the problem of inaccurate tail data in long-tail data sets,Seesaw Loss is introduced to balance the gradient of positive and negative samples by mitigating factor M and compensation factor C,and increase the accuracy of tail samples.In order to improve the accuracy of the regression prediction box,CIo U Loss is introduced into the regression loss to improve the regression accuracy of the model prediction box.Experiments show that the accuracy of Swin Transformer is 91.3%,and the accuracy of HESWTR is 95.6%,an improvement of 4.3%.Ablation experiments and comparison experiments are performed at the same time.The wild animal detection results show that the anchor box of the HESWTR network has higher confidence,more accurate recognition,and better model robustness.Figure 63 Table 14 Reference 79...