Research On Object Detection Based On Deep Learning

Object detection is a research hotspot in the field of computer vision.It refers to judging whether an object is included in the input image or video,and classifying and locating the detected object,which provides basic image or vdieo information data for other computer vision tasks.With the rapid development of deep learning algorithm based on convolutional neural network in recent years,object detection algorithm based on deep learning has made great progress.Compared with the traditional object detection algorithm,the object detection algorithm based on deep learning accuracy is higher,faster,and has become the mainstream algorithm in the field of object detection.However,there are still some defects such as low detection accuracy for small objects and low detection speed for real-time detection.First,this paper introduces the basic theory of object detection algorithm based on deep learning,which summarizes the main points of deep learning,the predecessor of CNN — multi-layer perceptron,and the basic principles and main components of the convolutional neural network,the classic convolutional neural networks and object detection algorithms.This part is the basis for subsequent improvement work.Second,this paper proposes an improved DB-YOLOv3 object detection algorithm based on YOLOv3 algorithm in view of the above defects in the object detection algorithm based on deep learning.DB-YOLOv3 algorithm proposed an improved Downsample Block based on Inception structure to improve the feature extraction network in YOLOv3.Downsample Block extracts features by means of convolutions and pooling of receptive fields with different sizes,which reduces the loss of feature information in the extraction process.Meanwhile,DB-YOLOv3 algorithm improves the reuse of deep features,which are high semantic information.DB-YOLOv3 algorithm effectively reduces the loss of feature information in the process of transmission,and meanwhile reduces the parameters of the model,which significantly improves the detection accuracy and speed of the model.Finally,in this paper,on the basis of the classical algorithm of SSD,aiming at the defects of SSD algorithm in detecting small objects,a DFF module based ondeconvolution and feature channel fusion is designed,and the detection network of SSD algorithm is improved by DFF module,which improves the reuse of deep and high semantic features in the detection network,and enhance the connection between shallow and high resolution features and deep and high semantic features.Experimental results show that the improved DFF-SSD algorithm significantly improves the detection effect of small targets.