An Improved Method Of Deformable Parts Based Model

Machine vision is an important branch of the artificial intelligence and object detection/recognition is an essential and challenge task of machine vision.Now,machine vision has been widely used in various research directions,such as manufacture,medicine,security,criminal investigation,atmosphere,traffic and military.However,the technology of object detection has been applied in various research fields in practice and promotes the development of each fields,the accuracy of object detection still depend on the quality of image samples.In practice,the quality of images shot by cameras are not ideal.For examples,there are different factors which may disrupt the images,such as noise,deformation,torsion,illumination and occlusion.Hence,the performance of object detection may decrease.Therefore,we propose a deep learning method based on deformable part based models to decrease the effect of these factors and improve the performance of object detection.The main research context of this paper are as follows:1.This paper designs a self-adaptive weights assigning system that assigns optimal weights to the two features in the combined features.The combined features have the ability of effectively representing the feature informtion of an object by combining the outline features and detailed features together.2.This paper designs a convolution neural network to extract deep features information of an image.These deep features are abstract features and have rich information relative to the hand-designed features(e.g.HOG features and SIFT features).And deep features are robust to the factors which may disrupt an image.Therefor,they are suitable for representing an object.3.This paper designs a gating architectural which can automatically select the perspective of an object in an image and can use the salience map to remove the noise form the back-ground.4.This paper designs a parallel deep net works which can learn the information of part deformation,occlusion and part feature level details.Since this architecture can learn multiple information simultaneously,hence the rich representation will improve the performance of detection.