Real-time Object Detection Based On Feature Enhancement

Real-time object detection based on convolutional neural networks is a key technology in the field of computer vision.The object detection based on convolutional neural network can be divided into two types: one-stage regression detection and two-stage detection.One-stage regression detection enables fast object detection,but the recall rate is not high,especially detecting small targets.Two-stage detection enables high-precision object position detection,but detection speed is slow.Moreover,many current target detection models have problems such as quantization error,loss of target features,and unbalanced classification.In response to the above problems,in this paper,we propose a real-time target detection framework based on feature enhancement.Our approach consists of five optimization options: Bilinear interpolation up-sampling,Convolutional network fully sharing,Light-weight position sensitive score map,Upward rounding quantization,Single fully connected layer.The bilinear interpolation up-sampling enhances the feature information of the target in the feature map,and the convolutional network fully sharing effectively avoiding the feature position error caused by mapping the target proposal region and the position sensitive score mapping feature map to the position feature.The design of light-weight position sensitive score map simplifies the network structure,eliminating the limitations of object type parameters,and solving the problem of unbalanced number of channels for classification and detection.The upward rounding quantization further enhances the feature information of the object,and the single fully connected layer makes full use of the enhanced target information,improving the accuracy and speed of target detection.We performed comprehensive experiments on a publicly available dataset DACSDC,Compared with R-FCN,our method can achieve an improvement of 8.5% on accuracy and a speedupof 1.4x on throughput.Our work ranked the 6th out of 24 teams in the 2018 system design contest on the 55 th Design Automatic Conference(DAC),and the achieved mean average precision(mAP)is 0.6317 with a throughput of 24.67 frame per second(FPS).