Real-time Target Detection Based On Regional Convolutional Neural Network

In recent years,object detection region has made remarkable achievements and has been continuously improved both in terms of speed and accuracy.However,existing object detection algorithms still have some shortcomings.For example,the algorithm usually trains multiple detectors for different scales of objects.In this way,the number of samples in the training set is distributed to each detector,thereby reducing the number of positive samples in the training process.On the other hand,the last non-maximal suppression method of the target detection algorithm directly ignores the non-maximal confidence detection result.In addition,because of the huge number of prior boxes produces by existing networks,it takes a lot of time to get the final results from all prior boxes.Based on the above shortcomings,this paper presents a variety of improved methods to ensure that the detection algorithm can be real-time operation,while further improving the efficiency of the network.The specific work of this paper is as follows:1)A scale invariant object detection network.In this paper,we first propose a new theory of invariant scale detection,and introduce this theory into the existing target detection network so that the same test results can be generated for the same objects at different scales.On the other hand,this paper proposes a new non-maximal weighting method.In order to obtain the only best target position among the repeated detection results,the non-maximal weighting method proposed in this paper integrates all the repeated detection results into one area by calculating the weighted mean to obtain the final detection result.2)A new matching strategy between prior boxes and the ground true object areas in detection algorithm.In order to improve the detection performance of the detection algorithm on small-scale target objects,a large number of prior boxes of small area are often generated,which also greatly slows down the speed of the algorithm.In this paper,combined with other target detection algorithms,small objects are matched by the block they are locate at,which greatly reduces the number of prior boxes of the algorithm needs to be generated,thus speeding up the algorithm speed.