General Object Detection Technology And Implementation Incorporating Face Detection

In recent years,with the continuous improvement of computing capacity and storage capacity of hardware,and the continuous innovation of communication technology,it is more and more convenient for people to obtain image data and transmit image data,but at the same time,the amount of image data in human society is also increasing rapidly.How to use computer to help human automatically identify and process massive image data has become a research hotspot.Object detection is one of the basic directions in computer vision.It provides the most basic information for computer vision applications.It is gradually applied in all walks of life,changing the mode of life and production.The task of object detection is to automatically locate specific visual object instances in a given image.Because the object detection algorithm based on deep learning has excellent detection performance,it has become a research hotspot in the direction of object detection.The object detection algorithm based on deep learning relies on strong supervision information to achieve high detection precision.In practical applications,if we want to expand the object category of the detection model,we need to label the dataset manually according to the object category.With the increase of the amount of data,it is more and more time-consuming and laborious to expand the object category.In this thesis,cross-dataset training technology is used to avoid the work of labeling datasets.While maintaining the precision of the model detection,the object category of the object detection model is expanded.The main contents of this thesis are as follows:(1)Firstly,two detection methods of object detection algorithm based on deep learning are introduced: anchor-based method and keypoint-based method,and the advantages and disadvantages of the two methods are discussed.The method based on the anchor adopts the dense anchor preset by human to acquire the positive and negative samples,and the method based on the keypoint adopts the feature map pixels to acquire the positive and negative samples.For different application scenarios,two kinds of object detection models with different emphasis are designed.The dual branch prediction model focuses on high detection precision and is suitable for GPU server computing.Lightweight RFBNet model focuses on high detection efficiency and lightweight,which is suitable for embedded(mobile)device computing.The dual branch prediction model combines the two detection methods to achieve the complementary advantages of the two detection methods.In view of the lack of high-level semantic information in the shallow feature extraction network,the receptive field module is introduced into the shallow network to enhance the high-level semantic in the shallow network features and further improve the detection precision.The lightweight RFBNet model uses the depthwise separable convolution instead of the ordinary convolution to compress the model,which improves the detection efficiency and compresses the model size.(2)For the shortcomings of the object detection model that requires additional manual annotation when expanding the object category,the cross-dataset training technology is used to train the general object detection dataset and face detection dataset jointly,and the dataset aware loss function is used to avoid the model classification confusion caused by the joint data training,so as to realize the general object detection model incorporating face detection.The experimental results show that compared with the single dataset training model,the detection precision of cross-dataset training model is almost the same,and the object category of detection model is expanded.