Research On Small Object Detection Method Based On Improved Convolutional Neural Network

The object detection task,as one of the basic tasks of computer vision,has been widely used in the fields of UAV aerial photography,remote sensing images,and unmanned vehicles.Currently,the detection effect of large objects in object detection has achieved a high level,while small objects have poor detection effect due to their low resolution,containing less feature information and easy to lose information in the process of convolutional neural network propagation.In order to solve the problems in the process of small object detection,this paper uses the SSD object detection network model as the basis for research,and improves the algorithm from the following three aspects:(1)A new bounding box regression loss function algorithm NIo U is proposed.A good regression loss function can optimize the network model parameters and enhance the feature extraction ability of convolutional neural network,which is very important for small object detection where features are lacking.The traditional bounding box regression loss function tends to lose the bounding box regression learning direction.In this paper,we strengthen the focus on multi-scale objects,and improve the regression algorithm based on the predicted bounding box and label bounding box centroid distance and overlap area ratio,adding a total of three types of penalty terms for the width and height scale loss as constraints.Through experiments,it is found that the loss function enables the model to regress correctly,improves the detection accuracy of the model for small objects,and effectively avoids the occurrence of phenomena that seriously affect the convergence accuracy of the object detection model,such as missed detection and false detection of objects to be detected.(2)An attention mechanism algorithm based on multi-scale channels is proposed.Since small object features have little information and easily lose some information in the process of convolutional neural network propagation,this paper adopts an attention mechanism to improve the attention of the model to small object features.To address the problem that general channel attention algorithms neglect spatial information acquisition,this paper first slices the channels,extracts the spatial information on each channel feature map with multiscale features,and then obtains channel attention weights from different scale feature maps.The algorithm promotes information exchange among channels,and the association of spatial and channel information strengthens the contribution of the overall object features to the final model detection results,and solves the problem of small object information loss during the neural network propagation.(3)An improved feature fusion network is proposed.Due to the small number of small object pixels,feature information is easily lost during the downsampling process.This paper designs a bottleneck module that integrates the attention mechanism to increase the interaction of object feature information and reduce the loss of object information.A feature fusion network that fuses high and low level features is proposed.The fusion network effectively combines the detail information of the low level feature maps of convolutional neural network and the semantic information of the high level feature maps,which greatly enhances the localization of small targets and improves the detection accuracy.In order to verify the performance of the improved method proposed in this paper in practical application scenarios,a small object detection system is designed and implemented to detect the mask wearing situation of people in the input image,visualize the detection results and record them in order to meet the demand for mask wearing detection in epidemic prevention and control situations.Finally,a functional test was conducted to verify the feasibility and effectiveness of the detection system designed in this paper.