Real-time Object Detection Using Deep Learning

Witnessing the advance of AI technique,object detection techniques have been widely employed in many commercial applications such as auto-drive,surveillance and robotics.Limited by the inefficient enumeration on objects and poor feature representation,the per-formance on the traditional approaches cannot meet the requirements in practice.Recently,with the succes of convolutional nerual network(CNN),the performance of object detec-tion has been significantly improved.Compared to the traditional approaches,the CNN models adopt an efficient strategy to predict objects,and adaptively extract the deep seman-tic features for different categories,thus could achieve an effective detection in real time.However,the current CNN approaches employ a single feature map to predict locations and categories of objects,it is prone to miss some tiny or overlapping objects.Besides,ignoring to learn the discrepant characters between categories,they may misclassify similar objects from different categories.To tackle these problems,our research aims to improve the performance of object detection from the three aspects:feature extraction,detector architecture,and optimization algorithm.In feature extraction,a gate-based filter is employed to integrate multiple raw features from the CNN.The gate-based filter could extract valuable information and block noises,thus it could provide a robust feature.Moreover,we design a two-branch detector to predict the presence of objects and categories respectively,which has a stronger learning ability as compared to the traditional CNN models using the one-branch detector.Finally,our optimization algorithm considers the discrepant information between categories,so the learned detector could better distinguish different categories.To verify the validity of the proposed approaches,we performed extensive experiments on two academic image datasets.The experimental results demonstrate that the gate-based filter,the two-branch detector and the optimization algorithm are all effective to improve the performance of object detection.Besides,our approach achieves a significant improvement as compared to the state-of-the-art methods.