Object Sketch Detection Model Based On Multi-layer Network

As one of the fundamental challenges in computer vision field,object detection attracts more and more attention from researchers.It is the critical step between image processing and image comprehension.Object detection aims to recognize targets from static images and dynamic videos,in order to complete the first step of image understanding and behavior analysis.Therefore,it is applied to various important fields intensively,such as video surveillance,intelligent transportation,biometric feature recognition,etc.However,object detection technology still has a long way to go.For the existing great range of target scales in real life scenarios,a single-scale model cannot balance the false positive rate and missing rate well.The solution of traditional detection methods is to construct an image pyramid,which always leads to the limitation of detecting efficiency.At the mean time,most detection methods focus on target localization.But for other fields of image segmentation,posture recognition of motion targets,what we need is not only a bounding box,but also more target descriptions and posture information.In addition,with the overwhelming research of deep neural network,new problems come into being,which are the necessity of numerous training samples and long training time.Take the convolved neural network as an example,the network model requires millions of training samples and dozens of hours for training at the least.All these obstacles have prevented further application of relevant theories.Confronted with the above problems,this paper proposes an object sketch detection model based on multi-layer network which stems from the improved active basis model.Active basis model can be learned from a few training samples and it is descriptive of the target so as to be beneficial for image understanding.However,active basis model aims to solve the problems of how to localize and describe the target.Furthermore,its learning step could be interpreted as a forward selecting process.Given a dictionary of Gabor filters,the original algorithm projects training samples to the dictionary through image convolution.Then it uses some maximumly pooled Gabor elements to represent the target by sparse coding.We introduce the notion of classification and explore the multi-target detecting problem instead of the single-target describing issue.Besides,the extension of multi-scale structure model has efficiently avoided the time consuming brought in by the image pyramid.We optimize the Pegasos algorithm and apply it to train the primary model again which contains the processes of the parameters optimization and the hard negative sample mining using a cluster of sample sets.As a result,the final model not only achieves high detection efficiency,but also has a more robust detection performance.The proposed model is implemented by a mixed programming of Matlab and C++,and it is tested on a self-established laboratory grape dataset and the public MIT pedestrian dataset.According to the experimental results,the proposed model has obviously surpassed the original active basis model on the detection efficiency and accuracy,confirming its feasibility and effectiveness.