Research On Technologies Of Object Detection Based On Visual Cognition

Based on the human visual cognition,this thesis carries out the research on technologies of object detection.The main research contents and contributions include:(1)Aiming at the problem that the existing saliency detection algorithms cannot suppress the background noise in saliency maps,a salient object detection method integrating background and foreground information is proposed.This method is inspired by the conclusions that the background information can highlight object regions and the foreground information can be used to suppress the noise.In order to obtain accurate foreground information of the image,the figure-ground mechanism of human visual system is further studied.By fusing contour features of the image in the process of region segmentation,a compact foreground region that conforms to the characteristics of visual perception can be obtained,from which the color features can be extracted to suppress the noise in saliency maps.Experimental results show that the F-measure of this method on ECSSD dataset is0.6887.It improves 6.1% compared with the SF method,and improves 2.5% compared with the LPS method.(2)Aiming at the problem of high complexity of natural images,and the poor performance of existing object proposal algorithms,an object proposal algorithm based on perceptual grouping is proposed,which is inspired by the researches of Gestalt Psycology.This method fuses the features inspired by the element connectedness cue in the existing framework to extract the object regions with diversified patterns from the natural image.Because the importance of grouping cues is quite different in organizing objects from different image scenes,a feature weight learning method based on multi-classifier is proposed.The training samples are first divided into groups to pretrain each individual model,and the algorithm adaptively adjusts the sample groups in the iteration procedure to optimize an ensemble of binary segmentation models.Experimental results show that at 1585 proposals,the proposed algorithm can achieve 67.3% average best overlap.It improves 3.8%compared with the Selective Search method,and improves 2.4% compared with the CPMC method.(3)Aiming at the inefficiency and low recognition rate of existing partial shape matching methods,a contour description method based on curvature information is proposed,which is inspired by the encoding and processing of shape contours in human vision system.The constructed descriptor is robust to the geometric transformation and local deformation of the shape contour.On the basis of the presented descriptor,a coarse to fine matching strategy is proposed according to the feedback regulation.The coarse matching stage aims to fast reject false candidates,while the fine matching is then implemented to verify the coarse results and correct mistakes.Experimental results show that in the task of MPEG-7 partial shape matching,the recognition rate of this method is 69%.It improves30% compared with the MVM method and improves 2% compared with the VDTW method.(4)Based on the above technologies of object detection,a ship detection method based on appearance and contour features is proposed.And a ship detection dataset is established to verify the effectiveness of this method.The detection results of the proposed method are robust to the background interference.In the experiment,the detection rate vs.false positives per image is used as evaluation index,and experimental results show that at 0.4 FPPI,the proposed method can achieve51.2% detection rate.It improves 21.2% and 37.3% compared with the traditional Structure-LBP and S-HOG method.Besides,a ship detection method based on saliency prior and deep learning is studied.It further improves the performance of the original model.