| Object detection is one of the core research contents in the field of computer vision,which plays a very key role in practical applications such as face recognition,intelligent transportation,intelligent agriculture,industrial detection and so on.The so-called object detection is to accurately locate and classify the target objects in the image.In reality,a large number of object detection methods have been proposed,which mainly include two categories: traditional object detection methods and deep learning-based object detection methods.However,it is noted that the current popular object detection methods are all general object detection methods,which may lead to the degradation of detection performance when directly applied to the detection of small objects.To this end,for the small object detection task,this paper takes the YOLOv5 deep neural network model as the benchmark,and proposes a new deep neural network small object detection method with attention mechanism.Here are the details:It is noted that the deep feature map of YOLOv5 deep neural network pays more attention to the extraction of small target features,and the spatial attention module is considered to be integrated into these deep feature maps to more accurately locate and identify small target objects.Firstly,the maximum pooling and average pooling operations are carried out along the channel axis of the feature map,and then the two are concatenated to generate an efficient feature descriptor.Finally,the fusion of multiple convolution kernels is used to generate a new spatial attention map for small object detection tasks.The proposed YOLOv5 object detection model with the attention mechanism can effectively extract the important feature information of multi-scale small objects,give more attention(weight)to small objects,and remove irrelevant redundant information,so as to improve the detection effect of small objects.Furthermore,in order to verify the effectiveness of the proposed object detection method,this paper collects a small flower object detection dataset containing 5248 images.The dataset contains a total of 20 flower categories,covering flower images with different lighting conditions,different flower densities,and different flower growth stages,and is divided into two types of flower targets: ultra-small targets and general small targets.The experimental results on this dataset show that the proposed deep neural network method with attention mechanism can pay more attention to the small object detection area,and effectively improve the small object detection ability.The F1 value on the ultra-small target dataset reaches 79.0%,and the m AP@0.5 value reaches 78.4%.It is 3.7% and 2.1% higher than the original YOLOv5 l model.The F1 value and m AP@0.5 value on the general small target data set are 0.9% and 1.6% higher than those of the original YOLOv5 s.In addition,in order to verify the generality of the proposed object detection method,experimental analysis is carried out on the TT100 k traffic sign small object detection dataset.The experimental results show that the F1 value of the proposed method reaches 88.4% and the m AP@0.5 value reaches 91.4% respectively,which are 2.4% and 2.6% higher than that of the original YOLOv5 x model.In addition,an example analysis is carried out.From the object detection results and the visualized image of the feature map,it can be found that the YOLOv5 object detection method with the attention mechanism pays more attention to the detection of small objects,and the proposed method can effectively improve the error detection and miss detection of the original YOLOv5 object detection method. |
