Research On Real-Time Generic Object Generator Of Super-Pixel Spanning Cascaded Multi-Threshold Extensions

The generic object generator,also known as the object proposal detection,has an ex-tremely significant role in accelerating the objects detection and enhancing the recognition accuracy.Because it can determine the possible position of the object in the image,narrow the search range of the traditional sliding window method and improve the recognition effi-ciency,the generic object generator has a very wide application in image recognition.The quality of the window set generated by the generator has a great impact on subsequent iden-tification.However,at present,most of the generic object generators are time-consuming or have a large positioning error.In some cases,it cannot meet the requirements of real-time processing.Therefore,this paper has carried out real-time general object generator research of super-pixel across multi-threshold cascade extension.First,we employ two generic object generators,BING and Edge Boxes,to generate the initial window set.The simulation results show that BING and Edge Boxes have the faster calculation speed and have higher recall rate under the lower positioning standard.Then,for the problem on parameter setting,low computational efficiency and a little poor segmentation effect of the graph-based segmentation method,for the first time,the adaptive threshold SLICO method is applied to the general object generator in this paper.Experiments show that this method can produce adaptive threshold for super pixel segmenta-tion,and has better segmentation effect at faster processing speed,which solves the above problems.Secondly,for the single-layer multi-threshold expansion(MTSE)method,a cascaded multi-threshold extension method is proposed which includes window focus calibration and cascaded extension.First,the window focus calibration is performed,that is,the windows generated by the initial generator is focused on the target body,and then the cascaded exten-sion is performed by global search and window diversification to improve the positioning accuracy of the window set.Experiments show that the cascaded multi-threshold extension method can effectively improve the window positioning accuracy.Finally,the window scoring and screening mechanism further screen the windows pro-duced by cascaded multi-threshold extension to upgrade the overall quality of the windows.Then,based on BING and Edge Boxes,two new generic object generators SCM-BING and SCM-EB are proposed.Based on the initial window set generated by BING and Edge Boxes,the image of LAB color space is firstly segmented by SLICO,and then the final window set is obtained by cascading multi-threshold extension method,scoring,filtering and sampling.By comparing with seven general object generators on the PASCAL VOC 2007 dataset,the experimental results show that SCM-BING and SCM-EB obtain the optimal window set at the expense of very little time loss,improve the accuracy of positioning,realize the balance of precision and time,and can be applied to real-time object recognition and other environ-ments.