Research On Object Detectio Method Based On Convolutional Neural Network

The task of object detection is to classify and locate a variable number of objects in image data.With the continuous progress of technology and the development of intelligence,object detection has become a difficult and hot issue in the field of computer vision,and has been widely used in intelligent classroom,automatic driving,unmanned supermarket,intelligent sorting of express delivery,intelligent medical and intelligent service robots.According to the types of methods used,object detection can be divided into traditional object detection methods and object detection methods based on convolutional neural network.Although the object detection method based on convolutional neural network has achieved better results than the traditional object detection method,there are still problems of location deviation,missed detection and false detection.Based on the research of object detection method based on convolutional neural network,this paper improves the clustering method and non-maximum suppression algorithm,so as to improve the positioning accuracy and detection accuracy of object detection model.The main work of this paper is as follows:Firstly,a new anchor clustering method,IOU-k-means++,is proposed to solve the problem of object location deviation.Through the research and analysis of yolov3 object detection model based on convolutional neural network,we know that in object location,using anchor to provide prior knowledge of object detection model location,we find that the anchor obtained by K-means clustering method in the original model has poor fitting performance for the labeled data,which makes the positioning accuracy of the object detection model poor.The algorithm uses K-means++,which is more accurate and stable in clustering,combined with the actual situation of object detection anchor calculation and clustering data type,and uses IOU distance to replace Euclidean distance as the measurement standard of clustering performance.The experimental results show that the anchor calculated by IOU-k-means++ has a higher fit for annotation data,and is more accurate for object positioning.Secondly,SF-NMS algorithm is proposed to solve the problem of missing and false detection.Through the research of non-maxima suppression algorithm in the post-processing stage of object detection,it is found that the greedy non maxima suppression algorithm is only used to delete the object frame directly through a single threshold when processing the repeated object frame,which has the problem of missing detection.Considering that the IOU is larger than the threshold value,there is also the possibility of belonging to the real object box.The algorithm combines the green NMS and soft NMS algorithm,and increases the double threshold to carry on the segmented punishment method,which well balances the missing and rechecking problems of NMS and soft NMS.Finally,this paper uses Pascal voc2012 data set to verify the algorithm proposed in this paper on the yolov3 object detection model,and analyzes the experimental results and the original model qualitatively and quantitatively.