Object Detection Method Based On The Contour Extraction In Image Segmentation

Multi-target detection and classification needs to detect and classify various targets out of the complex image-background,which is closely related to the target detection and image segmentation.At present,many new breakthroughs and several new research directions based on the Convolutional Neural Network(CNN)for the target detection have emerged,and hence a series of multi-target detection and classification methods corresponding to the CNN have constructed.In such methods,one category,which increases a Regional Proposal Network(RPN)to the convolutional neural network,is more accurate and more rapid than the others.Such category methods use such techniques as the regional convolutional neural networks,the Spatial Pyramid Pooling(SPP),and the fast-regional convolutional networks to make significant advances in multiple indicators of detection,while enabling the detection-speed to meet the realtime requirement,as well,the accuracy has been improved significantly.However,such kind of multi-target detection and classification methods are based on using the classification of the rectangular frame to mark the detected targets.Although the features on the speed and the accuracy have improved,the target position and shape contour still cannot be accurately calibrated in this kind of methods;Thus,for the purpose to increase the detection accuracy,the rectangular frame-based methods need a regression learning process on the bounding box approach,which increases the time cost of the entire convolution network.Besides that,the detection of small targets is not ideal.At present,the image segmentation is mainly based on the convolutional network to extract image features,on which the image segmentation is processed by adjusting the fully connected layer of the convolutional neural network,and finally different targets are represented by different pixels.However,such kind of methods is affected by the structure of the convolutional neural network(i.e.,there is a decrease in resolution during the convolution process,as well as the loss of receptive fields and target details,etc.),the segmentation accuracy is still not satisfied.Aiming at the above problems,In this paper,a method that extracting contour of target by image semantics segmentation is proposed to realize target detection and classification.Firstly,through the branch network,the proposed method obtains the regions of candidate based on the mapping relationship between the feature and the original image.Secondly,the qualified region be selected from the candidate region set,and its semantics segmentation map will be obtained by up-sampling and deconvolution.Furthermore,the proposed method uses a Full Connection Conditional Random Fields(CRFs)to optimize the segmented images,and to process the segmented images of the same class for obtaining the semantic segmentation maps of various targets.After that,the contour information and classification of the target has been obtained according to the semantic segmentation map.Finally,the contour information is used to label the target.Simulation results show that the proposed method not only can quickly detect and accurately mark the position and contour information of the target,but also can effectively improve the detection effect on the small targets with even better image segmentation accuracy.