Object Detection Based On Deep Convolutional Neural Networks

Object detection is an important task in the computer vision field,and it consists of two parts,object recognition and object locating.According to the pipeline of object detection,the solutions to the problems in the region proposal,feature extraction and object location are proposed.The specific research contents are as follows:1)Targeted at the problem that the image features extracted by current object detection algorithms based on convolutional neural network are not salient enough,we propose an object detection model based on an attentional region proposal network.We design an attentional region proposal net.An attention module is utilized to calculate the weights of the feature map,enhancing the salient features and weakening the background interference.Experiments show that the object detection models based on attentional region proposal networks improve the detection accuracy of salient objects without increasing the detection time cost much.2)To tackle the problem of insufficient object locating information in object detection,we design an object detection model with deep semantic information.The object detection model and the deep semantic segmentation model are combined to get the fused features that contain both object classification information and semantic information.Experimental results show that the proposed object detection model can locate the object more accurately and improve the accuracy of object detection.3)An object locating refinement network is designed to settle the problem of object locating accuracy.Search regions are expanded from the initial candidate boxes,and then are divided into five different parts which are sent to the locating refinement network.The networks are trained and tested by a single region or multiple region.It has been shown in experiments that the object location refinement networks we proposed have greatly improved the object detection performance.The way of expanding the candidate boxes and multiple different size of region help the network to locate the object better,leading to a more accurate position of the detection results.