Research On Object Detection Based On Mask R-CNN Model

With the rapid development of the times,more and more information fills the work and life of human beings,and the demand for intelligent acquisition of effective information from images or videos increases radically.As one of the important means to obtain information,target detection is widely used in video image analysis,regional search aerial photography,agricultural real-time detection,medical image research,vehicle safety supervision and other fields.This paper focuses on analyzing a kind of deep learning model in target detection,Mask R-CNN,and deeply analyzes the basic module and working mechanism of its general model.It combines theoretical knowledge with practice,and discussed the application of the model in specific situations.In this paper,target detection of power construction scenes is taken as an application example.Based on the Mask R-CNN general model,a multi-target detection system scheme for power construction scenes is designed,and the research is carried out on detection tasks such as helmet,personnel and text.The specific research work of this paper is as follows:(1)Basic modules and working mechanism of the Mask R-CNN model were analyzed,and a general model of Mask R-CNN target detection was established to obtain pre training weights.Combining theoretical knowledge with target detection requirements of power construction scenes,different transfer learning strategies were adopted according to different detection tasks,and parameter initialization was improved on the basis of pre-training weights for the Mask R-CNN backbone network.Aiming at two different detection tasks of helmet,personnel detection and text detection,a multi-objective detection system scheme for power construction scene is designed.(2)In the process of helmet and personnel detection,aiming at the problem of image targets being too small and different sizes in complex background,the multi-scale transformation method was used to adjust the regional suggestion network,obtain anchor points and conduct regression calculation to complete the detection experiment,and finally achieve the highest average accuracy of multi-targets.In the offline video detection module,aiming at the problem of poor video quality,the detection video is sharpened to achieve the purpose of image enhancement,and the accuracy is greatly improved,indicating that the structure of this paper strengthens the ability of feature extraction.(3)In the process of text detection,aiming at problems such as big power construction scenario text interval,attention mechanism is introduced to calculate and analyze significance in this paper.Combining channel attention with the characteristics of the pyramid structure constitute the pyramid channel attention mechanism,through the experiment proves that compared with the direct use of characteristics of the pyramid network,the design of pyramid channel attention mechanism has better characteristics.Indices like harmonic mean,precision and recall rate were improved.