Research On Deep Learning Based Single Shot Multibox Detector Optimization Algorithm

As one of the basic tasks of computer vision,object detection forms the basis for solving more complex or higher-level computer vision tasks such as segmentation,scene understanding and object tracking,and it is widely used in the fields of artificial intelligence and information technology,including robot vision,human computer interaction and automatic driving,etc.With the development of deep learning technology,the object detection algorithm based on deep learning technology has made great progress,Single Shot multibox Detector(SSD)is a object detection algorithm that provides the optimal trade-off among simplicity,speed and accuracy.This thesis chooses SSD as the base detector algorithm for research,considering the small object detection of this algorithm is not robust enough and the problems of samples and multi-task imbalances during the training process,this thesis made a profound study on how to improve the detection performance of SSD algorithm and proposed three effective solutions.The main contents of this thesis is as follows:Firstly,based on the detailed introdution of the network structure and principle of SSD algorithm,and in order to solve the problem of insufficient use of feature information due to the single utilization mode of detection layer in its network structure,a Two-way Feature fusion based Single Shot multibox Detector(TFSSD)is proposed.The TFSSD algorithm uses the proposed Two-way Feature Fusion Module(Tw FFM)to fuse the features of the conventional detection layers to generate new detection layers with rich geometric details and semantic information.The validity of the Tw FFM and TFSSD algorithm are proved by a series of comparative experiments on the public data set.Secondly,in order to obtain the feature information of the detection layers further,a Joint Attention Unit(JAU)is proposed.By embedding the JAU into the conventional detection layers,an Attention based Single Shot multibox Detector(ASSD)algorithm is proposed.The JAU consists of the Scaled Dot-Product Attention(SDPA)and the Squeeze-and-Excitation Block(SEB),and it can fully obtain more important and critical information and guide the model optimization by mining the correlation information in the detection layers from the two directions of space and channel.A series of experiments on public data set show that the JAU is effective,and the accuracy of ASSD algorithm is higher than that of conventional SSD algorithm.Finally,aiming at solving the problem of samples and multi-task imbalances during the training process,a More Balanced L1 Loss(MBL)is proposed.Furthermore,a Balanced with Two-way Feature fusion and Attention based Single Shot multibox Detector(BTFASSD)is proposed.BTFASSD algorithm uses the existing Tw FFM and JAU to build the overall network structure.BTFASSD algorithm first uses the Tw FFM to fuse the features of the conventional detection layers,and then uses the JAU to mine the key feature information of the detection layers,and last in order to achieve more balanced training the MBL promotes the regression of critical gradient by adjusting the weight of gradient contribution of easy and hard samples samples in the process of training.A series of comparative experiments on public data set show that the BTFASSD algorithm further improves the accuracy of SSD algorithm,especially the performance of small object detection.