Research On Dense Object Detection Based On Deep Learning

In recent years,great progress has been made in object detection.Under the conditions of severe occlusion,blur,and varied postures,better detection results can be obtained,but the detection of dense objects is still a challenging task.Therefore,this thesis develops and analyzes the dense object detection algorithm based on deep learning.Finally,the validity and rationality of the proposed algorithm are verified by dense course name and dense face.The main research contents are as follows:In the case where the dense object distribution is fixed in the image,this thesis verifies and analyzes the problem that the detection result of the existing general object detection model is not good when it is directly used for dense object.In order to improve the detection accuracy of dense object,a fixed distribution dense object detection algorithm based on deep learning is designed.The dense course names in the transcript are studied as an example.In order to reduce the intensity of the object and avoid the phenomenon of object missed detection after the segmentation,an image overlap block method is designed according to the object distribution rule.The block is detected using the YOLOv2 model.After the block detection,if there is a duplicate object,an appropriate method is selected to remove the repeating boxes.In addition,if there is a tilt angle of the object,the tilt object is corrected to improve the detection accuracy of the object.In order to verify the performance of the algorithm,the experiment is carried out on the transcript data set.The experimental result shows that the detection accuracy of the algorithm is 99.11%,and the missed detection rate and false detection rate are only 0.89% and 0.89% respectively,which has the best detection result and is better than YOLOv2 and Faster R-CNN.The above algorithm can only solve the fixed-distribution dense object detection problem.Therefore,for the dense object with unfixed distribution,this thesis designs a variable distribution dense object detection algorithm based on YOLOv2.First,the algorithm defines a density formula that correctly reflects the object intensity and designs a density threshold as the object density criterion.The image needs to be overlapped and segmented when it contains dense object.After the block images are detected by the YOLOv2 detection model,the object location is returned to the original input image.If there are repeating boxes,the fusion method is used to fuse the repeated boxes to obtain the final detection result.In addition,the object annotation information in the training set is clustered,and the width and height parameters of the object candidate bounding box are re-selected,and then the model is trained to improve the positioning accuracy.To verify the validity of the algorithm,experiments are performed on dense face data set and FDDB face detection data set.The experimental results show that the detection accuracy of the proposed algorithm on the dense face data set reaches 96.75%,and the missed detection rate and false detection rate are 3.25% and 3.14% respectively.Compared with YOLOv2 and Faster R-CNN,this algorithm not only improves the accuracy of dense face detection,but also reduces the rate of missed detection and false detection.At the same time,it also has better face detection effect on FDDB data set.Compared with the face detection algorithm SSH,this algorithm reduces the false detection rate.In order to show the algorithm more intuitively,this thesis designs a demonstration system to demonstrate the functions of each step of the algorithm.