Research And Application Of Multiscale Target Detection Based On Full Convolutional Network

Object detection technology is often used to locate objects in images,this technology has great application value in industry,has been widely used in intelligent monitoring,human-computer interaction,national security protection and other fields.In order to meet the various needs of human beings,the target detection algorithm is also improving.Traditional object detection methods often build target features by man,use the method of sliding window to obtain regions of interest,and establish classification models by manual selection of classifiers for regional classification.There are two main problems in the whole classification process of the model: on the one hand,the process of traversing the image to obtain the area of interest is too time-consuming,and at the same time produces a large number of unrelated areas to be examined,which brings a lot of computation to the model for regional classification,on the other hand,the characteristic robustness of the hand-built network is poor,and the deep neural network represented by the convolving network has made amazing achievements in the target detection with its powerful characterization ability,and the goal detection based on deep learning has been paid attention and gradually become a hot topic of research.Deep learning target detection algorithm can be divided into two-stage series and one-stage series,the two-stage algorithm is highly accurate but the detection is too time-consuming,while the one-stage algorithm improves the detection speed while maintaining a certain accuracy rate,meets the needs of real-time detection,and is more suitable for practical application.In deep learning,the YOLO series is one of the representative algorithms in the one-stage target detection algorithm.In this framework,a simple convolution module is applied to extract the target feature,and the up and down sampling technique,skip layer and other operations are also introduced to obtain multi-scale Feature map and multi-scale prediction.Based on YOLO v3 network,this paper studies the problem of laboratory test and standard dress for laboratory mice,and discusses the implementation principle of the target detection framework of the whole convolution network.In the target detection of laboratory mice,because the appearance and background of the experimental mice are relatively single,the lightweight detection network-YOLO v3-tiny algorithm was used to detect the target of the experimental mice.For the test of the experimenter's standard dress,taking into account the light change and the noise influence of the laboratory door,the YOLO v3 with deeper convolution is selected as the detection network.The main research contents of this paper are as follows:(1)We study and implement the mice detection approach as a traditional computer vision topic.In order to solve the problem that the traditional detection method based on image threshold segmentation is susceptible to light change,noise and other influences resulting in poor detection effect,first observe the characteristics of the experimental mice,and construct the features from the gradient amplitude and gradient direction according to the law of image texture smoothness and uniform background gradient,and then use the histogram to reduce the feature dimension and combine the SVM(support vector machine)classifier to realize the new detection method of the experimental mice.The experimental results show that the precision and stability of the new method are significantly improved and more applicable than the image threshold segmentation method.(2)YOLO v3-tiny network is applied to realize multi-scale detection and optimization algorithm for laboratory mice.In this paper,the actual collection of experimental mouse data as a training sample,first of all,analysis of the size of the experimental mice in the training set,determine the network training parameters and optimize the number of anchor points of the a priori box,in the detection process through pre-processing to enhance the image characteristics.The YOLO v3-tiny was trained using the Pytorch deep learning framework under the Win10 system and evaluated the performance of the optimized network using mAP values.Experiments show that YOLO v3-tiny,which increases multi-scale prediction,is more accurate and has a shorter running time than the traditional SVM classification model,and its optimized algorithm mAP value is higher.(3)We build a YOLO v3 network to monitor the dress of the experimenter.First of all,the monitoring video at the laboratory gate is used by labelImg software as a training set,and then the Darknet framework is built,and then the height ratio of the experimental staff is analyzed,the network parameters of YOLO v3 are determined,and the accuracy of the algorithm is improved in order to reduce the interference of light and noise in the prediction level of YOLO v3,and finally,the feasibility algorithm is improved by making experimental comparison.