Research On Multiscale Detection Based On Convolutional Neural Networks

In recent years,the field of artificial intelligence has developed rapidly.Especially in the field of object detection,a large number of detection models based on convolutional neural networks have been proposed.Although these models have achieved good results,they have not worked very well in multi-scale object detection.Thus,it is of great significance to construct a model that can detect objects with multiple scales in the field of target detection.Therefore,we proposes a multi-scale faster region-based convolutional neural network(MSF-RCNN)model.The model includes two sub-networks: region proposal sub-network and classification detection sub-network.The region proposal sub-network outputs the feature map corresponding to multiple convolution layers to expand the scale range of object detection.After that,the anchor algorithm is used to further expand the object detection scale range,thereby achieving the purpose of detecting multiple scales of the object.The classification detection sub-network includes two full-connection layers and two-channel small full-connection layers.This structure can increase detectable object types and achieve the purpose of detecting multiple types of objects.The deconvolution layer is used to connect between the two sub-networks.The convolution feature map received by the classification detection sub-network is enlarged and the resolution of the map is improved,and the deconvolution layer is fixed to avoid the influence of the two sub-networks.In short,the MSF-RCNN model is an end-to-end multi-objective detection model based on multi-task loss function,which can detect objects of various scales at the same time and ensure good accuracy.By comparing experiments on the voc dataset,the MSF-RCNN model achieved a detection speed of 10~11 frames per second,and it performed well in all indicators such as recall,precision,miss rate,and false positive per image.The scale of the detectable object is wider,and the performance of the MSF-RCNN on the coco dataset is equally good.What's more,by comparing the experimental results of MSF-RCNN-2x with deconvolution layer and MSF-RCNN without deconvolution layer,it can be seen that the resolution of the convolution map can be improved by the deconvolution layer,which has a positive effect on model performance.