Research On Small Object Detection Algorithm Based On Attention Mechanism

The current object detection algorithms have good performance on large and medium objects.However,for small objects in different scenarios,they do not balance the detection accuracy,speed and model complexity well.Aiming at different types of datasets containing small objects,we introduce the attention mechanism into the appropriate location of the network and combine the lightweight idea and other methods to study and propose three small object detection algorithms based on attention mechanism,which effectively balance the effectiveness and efficiency of the algorithms to a certain extent.(1)Aiming at the lower detection accuracy of YOLOv3-tiny on the self-built safety helmet dataset,SAS-YOLOv3-tiny based on attention mechanism is proposed in this thesis.The convolution layer and lightweight Sandglass-Residual module based on channel attention are used to reconstruct the backbone network to simplify the network model and obtain more expressive features.The three-scale feature prediction method is utilized to effectively integrate the features of small objects.The improved Spatial Pyramid Pooling(SPP)module is added to the back of the backbone network to extract local and global features.Finally,CIo U is used to improve the positioning accuracy of small object.Experiments show that SAS-YOLOv3-tiny improves the detection accuracy while reducing the amount of model parameters,and it’s lighter than other models.(2)Aiming at the higher model complexity,slower detection speed and lower detection accuracy of YOLOv3 on KITTI and CCTSDB datasets,YOLO-MXANet based on attention mechanism is proposed in this thesis.The SA-Mobile Ne Xt based on the Shuffle Channel and Spatial Attention(SCSA)module is proposed to reduce the complexity of model while not weakening the ability of feature extraction as much as possible,and the Multi-scale Feature Enhancement Fusion network is used to better fuse the features.Moreover,the data enhancement method combined with Mosaic and Mixup is used to promote the quality of the datasets,and Si LU activation function is employed to accelerate the convergence of training model.The Experiments show that when the input is640?640?3,the number of parameters and calculation cost of YOLO-MXANet are reduced by4.763?10~7 and 117.9 GFLOPS respectively,the detection speed is increased by0.6ms,and the m AP value is increased by 1.3 percentage points compared with the baseline on the KITTI dataset.It also has better detection effect compared with other lightweight algorithms.(3)Aiming at the lower detection accuracy of YOLOv5s(v 6.0)on PASCAL VOC dataset,YOLOv5s-ANG based on attention mechanism is proposed in this thesis.The v5s-Backbone-A based on attention mechanism is proposed to obtain the key features of common objects.In addition,the combination of K-means clustering and genetic algorithm can obtain anchor boxes that better match VOC dataset,and the Ghost convolution is used to reduce redundant features.Experiments show that the m AP value of YOLOv5s-ANG is improved by 3.4 percentage points,and the number of parameters is reduced by0.510?10~6.The extended experiments show that SAS-YOLOv3-tiny and YOLO-MXANet have certain generalization on PASCAL VOC,and YOLOv5s-ANG has the better detection performance.In conclusion,the three small object detection algorithms based on attention mechanism simplify the model by using lightweight ideas and other methods,while enhancing the features of small objects including safety helmets,pedestrians,vehicles,traffic signs and general categories.