Depth DPM Algorithm And Its Application In Object Detection

Object detection is an important research direction in the field of computer vision,and plays an important role in all walks of life.Object detection uses different machine learning methods to solve various problems in object detection,but it still has various challenges.Convolutional Neural Networks(CNN)is a nonlinear classifier for "black box" that can automatically obtain advanced features of images through their hierarchical structure.Despite this,CNN does not provide a clear relationship between lower-level features.Therefore,it may lose potential information about the candidate relationship structure,which is important information for improving accuracy.The Deformable Parts Model(DPM)algorithm is a graphical model(Markov random field)that uses a series of parts and the spatial positional relationship of parts to represent an object.In the object detection,both DPM and CNN have their own advantages and disadvantages.Therefore,integrating DPM and CNN is a promising method.At the same time,the objectness measure algorithm can provide fewer proposal windows than the sliding window search strategy,reducing the computational complexity.Therefore,this thesis focuses on how to combine the objectness measure algorithm,the CNN algorithm,and the DPM algorithm to give full play to the advantages of the three,and improve s the detection efficiency and detection accuracy.(1)The BING algorithm is currently the fastest detection algorithm for objectness measure algorithm,which BING has a running speed of 300 fps.The recall rate is 96.2% on the premise that the intersection-over-union threshold is 0.5 and the first 1000 proposal windows are taken.However,there is a problem with the BING algorithm.With the growth of the intersection-over-union threshold,the detection performance of the BING algorithm will drop rapidly.In the second chapter,by analysis of the causes of success and failure of the BING algorithm,this thesis clarifies the correlation between its detection effect and the intersection-over-union threshold,so as to obtain the relationship between parameters of the BING algorithm,and sets parameters to improve detection performance.(2)The DPM algorithm uses a sliding window strategy to search for the object on an image,which is an important factor that restricts the detection efficiency of the DPM algorithm.In the third chapter,this thesis uses parameter-selected BING algorithm instead of sliding window search strategy to provide proposal windows,which can effectively reduce some redundant windows.In order to use the BING algorithm and the DPM algorithm jointly,this thesis proposes a mapping formula between a proposal window and a feature map,and sets different search ranges according to the diffe rence of the root filter and parts filter.Through the joint use with the BING algorithm,DPM's detection efficiency has been improved.(3)In the fourth chapter,aiming at the situation that the Deep DPM calculation feature pyramid is too expensive,this thesis obtains the scale invariance of the algorithm by adding normalized layer and using multiple DPM-CNN with different filter kernel size instead of feature pyramid,keeping the robustness of the scale changes while reducing the computational complexity.At the same time,this thesis uses BING algorithm to get proposal windows,to reduce the number of search windows.It achieves the purpose of using BING,CNN and DPM algorithm jointly,exert ing the advantages of the three,and improving the detection efficiency and performance.