Small Target Detection Based On Deep Learning

The detection of targets in images or videos has always been an important research content of computer vision technology.Its main task is to determine whether the input image or video contains the target to be detected;to mark the exact location of the target,and at the same time,it is necessary to determine the category of different targets in images or videos,among them,the detection of small targets has greatly practical application value in many technical fields,such as photoelectric detection,automatic driving and biomedicine.With the advancement of modern digital imaging processing technology and data storage technology,people can easily obtain massive data in various forms from images.At the same time,with the popularity of deep learning,the processing ability of this technology in the field of vision has begun to show great advantages over traditional methods.However,because the proportion of small objects in the image is small and the features are not obvious,compared with the conventional target in engineering,deep learning method also has the problems of high miss detection rate and low accuracy rate when detecting small targets.Therefore,in view of the above defects,this paper carried out the following researches:First of all,this paper sorts out the current commonly used target detection algorithm model based on deep learning,analyzes the main advantages and disadvantages of each target detection algorithm,according to the basic requirements of target detection for real-time and accuracy,and finally determines to design a small target detection model on the basis of YOLOv3 convolutional neural network.Secondly,using tensorflow to implement YOLOv3,and set up a set of small target image data set for training,verification and testing of small target detection model.Secondly,aiming at the problem that the YOLOv3 convolutional neural network model has insufficient detection capabilities for small targets,the network structure of YOLOv3 is improved,and the detection capability of the network model for small targets is improved by increasing the depth of the network,achieving dense connections and expanding the size of the detection feature map.After comparative analysis of experiments,the improved network has improved the detection ability of small targets.At the same time,the paper also tests the improved YOLOv3's ability to detect the target in the infrared image,and the test results are compared with the test results of visible image.Finally,for the improved YOLOv3 model,there may still be a small number of small targets miss.Combined with the fact that the detection results of different small targets in the visible image and infrared image are different,this paper designed an improved YOLOv3 small target detection algorithm with multi-frame fusion,the detection algorithm selects the appropriate fusion threshold,and the detection results in the visible image and the infrared image are merged and displayed in the visible image,and the experimental results show that the algorithm is effective.