Research On Small Target Detection Based On Improved Convolutional Neural Network

As one of the important research directions in the field of computer vision,the goal of object detection is to use corresponding detection methods to identify whether a specific target is present in an image.Currently,deep learning-based object detection algorithms have made rapid progress in various fields and become a research hotspot.However,small objects contain less information,have low resolution,and are prone to information loss during the propagation process of convolutional neural networks,which affects the detection results.Therefore,methods for small object detection have become a research hotspot,which is of great significance for fields such as aerial photography and drone monitoring.Since R.Joseph et al.proposed YOLO in 2015,YOLO series models have been widely praised for their fast detection speed and high accuracy,and have become one of the most outstanding and versatile models in recent years,showing good performance in small target detection.In this paper,we choose YOLOv5 as the baseline network and improve and optimize it to enhance its small target detection capability.The main contributions of this paper are as follows:(1)Changing the Backbone structure used for feature extraction in the YOLOv5 baseline network,and testing the impact of different Backbone structures on the speed and accuracy of small target detection.Small targets often contain less feature information due to their inherent characteristics.As an important part of the object detection network for feature extraction,different Backbones extract and calculate feature information in different ways.A suitable Backbone feature extraction network can retain more feature information without losing detection speed,which is conducive to improving the accuracy of small target detection.(2)Introducing the ECA attention mechanism into the Bottleneck CSP module to overcome the contradiction between performance and complexity,which can make the network better focus on the regions of interest in the image and enhance the performance of small target detection in complex backgrounds without increasing the complexity of the network.(3)In order to reduce the number of parameters,reduce the time problem caused by unnecessary calculations,this paper introduces the Ghost Conv in Ghost Net to replace the general convolution method in the main feature extraction network.In the process of small target detection,due to its own characteristics,there are many redundant calculations,and Ghost Conv is conducive to eliminating redundancy,reducing the number of parameters,and accelerating convolution speed.(4)The upsampling method used in YOLOv5 did not fully consider the characteristics of small targets and would lose a lot of feature information during the sampling process.This paper introduces transpose convolution to replace the sampling method.Transpose convolution can set automatically learning parameters,and better retain target feature information during upsampling,which is conducive to combining shallow and deep feature maps.(5)The initial model of YOLOv5 sets three detection heads and corresponding to them,sets three anchors.The feature map is divided into three channels of different sizes,with a total of nine different anchors.In order to better detect those targets that are too small,the new model adds an additional small target detection head,and sets smaller anchors,which can effectively frame small targets.To verify the application of the proposed method in actual scenarios,this paper selects a drone aerial dataset for testing,detects various targets in the shooting scene,and calculates the accuracy.Under the YOLOv5 l baseline model,the improved model’s m AP50 has increased by 10.3%,m AP50:95 has increased by 7.4%,and the model’s m AP50 can reach 62.8% when trained at an input resolution of 1536*1536.