Research On Object Detection With Deep Learning

Object detection is to determine the targets' location and category in one image or video sequence.It is a research hotspot in computer vision and pattern recognition in recent years and becoming more and more common in industrial applications such as intelligent monitoring,face recognition and pedestrian detection.A number of excellent new object detection algorithms have emerged with the rise of deep learning and its deep apply in the field of computer vision.These algorithms have outstanding performance in detecting accuracy or its real-time speed.This thesis focus on the object detection algorithms and its related field like deep learning,which deep apply in object detection.Then,YOLO and SSD,have made the improvement.The major contributions of the thesis are:(1)Summarized the basic operations and components of convolutional neural network.The basis of the CNN include the layers,the activation function,the objective function and the regularization and so on.The Darknet and VGG classification model,which are introduced in the thesis,are the foundation of the YOLO and SSD.Commonly,the detection frameworks are fine tuned from the classification model s.(2)Proposed an algorithm that add the Batch re-normalization to the YOLO network structure.It combining the characteristic of batch re-normalization,which has the strength to deal with small or non-i.i.d.minibatches.The improved algorithm view the feature maps,which generated from the convolutional layers,as activations,and then batch renormalize the activations,meanwhile it removed the Dropout from the original network structure and increased learning rate.The experimental results shows that the proposal algorithm has better detection accuracy and faster than before,and furthermore,it can decrease the model training time and the requirement of hardware equipment.(3)Proposed an Atrous filter design strategy,which can strengthen the resolution of feature maps.The improved algorithm concatenated the feature maps that generated by the third and fourth convolution layer after normalization,and then improves the resolution of these feature maps by Atrous computed.The concatenated feature maps provide the required features for small objects.In addition,the SSD improved algorithm also add Se LU(Scaled Exponential Linear Units)activation function and designed a data augmented methods in the data preprocessing phase.The experimental results shows that the proposal algorithm has higher detection accuracy and better robustness than the original SSD algorithm.Furthermore,the detection performance obvious better on small target detection.