Research And Application On Object Detection Based On Weakly Supervised Learning

Object detection based on fully supervised learning relies heavily on complete annotation data in realistic scenarios,making it difficult to apply traditional fully supervised learning methods in some realistic detection scenarios.In the weakly supervised object detection,due to the lack of location labeling information,the model tends to converge the detection results of objects,especially non-rigid objects,to the most discriminative local regions of the object,which leads to incomplete detection results of the object.In addition,the pseudo-labeling process focuses too much on the local region of the object with the highest classification confidence,which makes the mining of other positive examples insufficient.Therefore,this dissertation investigates weakly supervised object detection to address the above problems,and finally applies the research results to the field of wildlife detection.The research work and innovation points of this paper are reflected in:(1)To address the problem that weakly supervised detection networks tend to make the detection results of objects incomplete,a weakly supervised object detection framework based on double attention erasure and attention information aggregation is proposed,which contains two modules of double attention erasure and attention information aggregation.The double attention erasure module can effectively extend the most discriminative region and add the attention information aggregation module to further improve the detection accuracy.Experimental results on the datasets VOC 2007 and VOC 2012 demonstrate that the proposed method effectively extends the most discriminative region of the object and thus effectively detects a larger region of the object.(2)To address the problem of not being able to fully explore other positive instances when performing pseudo-label labeling,a weakly supervised object detection algorithm based on iterative labeling learning is proposed,which sets a threshold based on the highest confidence score and iteratively selects positive instances and labels them to explore more positive instances.In addition,a dense online instance classifier optimization algorithm is used to solve the information loss problem in the supervised information generation process and further improve the model accuracy.Experimental analysis on the datasets VOC 2007 and VOC 2012 shows that the proposed method achieves significant results.(3)A weakly supervised object detection system for wildlife detection is proposed to address the problem of high annotation cost and limited annotation quality of annotators in the field of wildlife detection.The system inputs the wildlife images to be detected into the trained model to obtain the final detection frame,and displays the detection results through a visual interface.The final demonstration shows that the system can accurately detect wildlife images.