Multi-scale Target Detection Research Based On Convolution Neural Network

With nearly ten years of development,machine learning has gradually become the main method of target detection.And convolution neural network as an important algorithm in machine learning has been widely concerned.Convolution neural network is a special artificial neural network.It is a new type of network combining the artificial neural network and convolution operation.Convolution neural network has the characteristics of sparse connection and weight sharing,which makes the number of network parameters greatly reduced.Convolution neural networks show good results in the identification of various targets and on the handling of warppings and distortions.Therefore,many researchers have done a lot of research and experiments on the convolution neural network,and put forward many detection methods and models based on convolution neural network which have obtained very good test results.Although convolution neural networks can solve many types of problems,more research is needed to detect objects on multiple scales.The convolution neural network for multi-scale target detection based on region proposal can detect targets on different scales in the image and improves the detection speed.The network consists Both of the two sub-networks are learn by end-to-end.The region proposal sub-network combines MSCNN's multi-scale strategy and RPN method.The sub-network uses some output layers to detect candidate regions,and each output layer concentrates on objects within certain scale ranges.The candidate region detectors on the different output layers are combined to form a strong multi-scale detector to obtain the final candidate regions.The target detection sub-network adopts the candidate regions obtained by the region proposal sub-network and the CNN feature maps as input.Then the final classification and positioning results are obtained by softmax classification and bounding box regression after ROI pooling.In the target detection sub-network,feature upsampling is used to approximates the CNN feature maps instead of input image upsampling.The deconvolution layer is added to increase the resolution of the feature maps so that small objects can produce a strong response area.The experiment results show that the network can effectively detect the targets of different scales in the image andrealize the function of multi-scale target detection.And comparing with the existing target detection methods,the network is shown a higher accuracy and faster speed.For the SSD model detecting small-scale objects badly,we proposals the convolution neural network for multi-scale target detection based on regression.Because the low-level network preserves the details of the image,the network improves the SSD model by adding the low-level CNN feature map information to detect the target,which is easy to detecting small-scale targets.In the training of the network,different data enhancement schemes were used to improve the training efficiency.Experiments were carried out on the network using the data sets of PACAL VOC2012 and MS COCO.The effect of data enhancement scheme on target detection was analyzed.The network model was compared with SSD and YOLO,and it is proved the effectiveness of the network.