Research On Small Target Detection Algorithm Based On Statistical Features And Bridge Method

Target detection is a key technology in the field of computer vision such as automatic identification,warning,and intelligent transportation.At present,detection algorithms are mainly divided into three categories: traditional methods,traditional machine learning methods,and deep learning methods.Since each of the three methods has advantages and disadvantages,the detection method of the specific target needs to be determined according to factors such as the degree of its characteristics and the complexity of the scene.In order to detect the target more effectively,the following work is done in this paper.1.Since the characteristics of large-scale fishing vessels in military ships and civilian vessels are very similar,the detection algorithms based on traditional methods or traditional machine learning methods are generally correct.In order to detect warships/civil vessels more effectively,this paper creates a deep learning-based detection algorithm.It mainly detects ship targets by image grid prediction and postGaussian processing.The main framework of the model is as follows: the image is meshed,and then the class probability is calculated separately.At this time,the local distribution of the output matrix is uniformly distributed.The data is then Gaussian processed such that the matrix is locally Gaussian-like.Finally,the target recognition and image localization are realized by detecting the Gaussian distribution or the Gaussian highest point.The experimental results show that the average accuracy rate of the deep learning network model in this paper is 1.4% higher than that of YOLO v2 network and SSD300 network,respectively,more than 2%,and the real-time performance of YOLO v2 network is increased by more than 1.5%.2.Weak and small targets have the characteristics of less pixel number,less obvious texture features,less shape information,and stronger local contrast characteristics.In order to detect weak targets more effectively,this paper creates a detection model based on traditional methods.This model mainly focuses on detecting weak targets.It mainly detects small targets by the feature difference between weak and small targets in the image and their neighborhood background.The main steps of the traditional model are as follows: Firstly,the mean value and variance of the pixel values are extracted within the sliding window;secondly,according to these statistical features and the bridge method,whether there is an small target in the window range is judged.If there is a small target,record its position;finally,the small target area is subjected to secondary screening.The secondary screening method is to select one of the screening methods based on statistical features,based on local extremum and based on connected domain shape.The experimental results show that compared with the more classical algorithms,the false alarm rate is reduced by more than 58%.3.At present,the detection of small targets is best with traditional algorithms,and the target with normal size is better with deep learning.This paper uses target size as the switching condition of the two models to detect the target.When the target size is less than the critical value,a small target detection algorithm is used.When the target size is greater than the threshold,a deep learning detection algorithm is used.