Research On Small Object Detection Method Based On Deep Learning

Small object detection is an important problem in the field of computer vision object detection and has a wide range of applications.However,small objects usually have characteristics such as low contrast,low resolution,and complex background,which makes existing object detection methods ineffective in dealing with small objects,so it is of great significance to study how to detect small objects efficiently and accurately.Improvement by using deep learning technology can effectively improve the accuracy of small target detection,which will promote the further development of small target detection technology and provide better results and experience for practical applications.This thesis studies the problem of small target detection based on deep learning.First,the basic concepts and common models of target detection are introduced,focusing on the analysis of the two-stage target detector Faster R-CNN and the single-stage target detector YOLOv5.Second,the two models are improved separately.Finally,the effectiveness and superiority of the proposed method are verified by comparative analysis of experimental results.The main work of this thesis is as follows:For the two-stage target detector Faster R-CNN,in order to solve the problem of difficult feature extraction of small targets,ResNet50 and VGG16 were used as the skeleton network for Faster R-CNN in the research process.Based on the original model,an improved IoU algorithm is used to solve the problem of imbalance between positive and negative samples,and an FPN feature fusion network is introduced to improve the multiscale detection ability of the detector.During the experiment,the improved Faster R-CNN model was trained using the transfer learning method,and the performance of the model in small target detection was evaluated.The experimental results show that compared with the original model,under the architecture based on ResNet50 and VGG16,the improved Faster R-CNN improves the mAP by 9% and 7.4%,respectively,which shows that the improved method can effectively improve the small target detection performance of the model.For the single-stage target detector YOLOv5,the model has a better recognition effect when the target is large,but for small target detection,it is prone to missed detection and false detection.Therefore,the SPD module is added to the model to improve the detection rate of small targets.SPD expands the receptive field of the convolution kernel,enabling the network to better capture the detailed information in the image.In addition,the CBAM mechanism is added.CBAM can adaptively calculate the attention weight of the channel and space,so as to extract important feature information in the image and better distinguish the target from the background.The improved model is trained and tested on the Vis Drone dataset.The experimental results show that compared with the original YOLOv5 model,the SPD and CBAM methods can better focus on useful features,thereby improving the detection accuracy of small targets.Compared with the original YOLOv5 model,the mAP index has increased by 4.6%.