Research On Object Detection Based On Hybrid Model

Recently,many researchers focus on computer vision,especially image object detection.The rapid development of deep learning technology provides a fertile soil for object detection and also indicates new research trends,various models based on convolutional neural networks continuously refresh the record of average precision.To avoid the weakness of the traditional machine learning method using the sliding window for enumeration,almost all the best perform object detection networks currently employ region proposals to guide the search of instances.Fast rcnn model is a typical and outstanding example employing region proposals.It decomposes the detection task into two steps.First,it extracts approximately 2000 potential object regions from the input image,and then those potential object regions are classified into different categories and located accurately.Although dealing with two relative simple tasks,it still has rooms for improvement.In this dissertation,we focus on getting better performance and optimizing the shortcomings of the fast rcnn model to further improve the average precision.Firstly,we trained a foreground/background binary classifier based on SVM to post-process the potential object regions,aiming to remove most of the simple regions while filter out those high quality regions.Experiments show that the classifier in this dissertation helps to lower the false positive rate of the objects classify.Secondly,unlike the fast rcnn that uses only the top feature in object detection,we merge highly resolution low-level features,highly semantic depth-level features,and complementary middle-level features into one strong discriminative feature map.Experiments show that the fusion feature greatly improves the performance and average precision of small objects.In addition,we replace the cross-entropy loss with focal loss to optimize the original model.The experiment shows that the focal loss function plays a similar role as hard example mining,making the network training more stable and robust.Finally,all the improvement and optimization measures are combined to train our hybrid Fast rcnn++ model which combines traditional machine learning and deep learning proposed in the dissertation.The results of experiments show that the model of us increases the object assessment index mAP from 67.2% to 72.2%,and compared with some models that are currently superior,we achieved competitive and comparable results.