Research On Small Object Detection Algorithm Based On Feature Enhancement And Attention Mechanism

As the basic research of computer vision,object detection is to classify and accurately locate objects appearing in images or videos.Compared with traditional detection methods,convolutional neural network based on deep learning has stronger feature extraction ability,and has greatly improving the detection speed and accuracy,but problems of false detection and missing detection still exist.Because small targets occupy small pixels,it is more difficult to detect small targets.Based on the object detection algorithm SSD,three improvements are proposed in this paper,including the addition of an improved feature fusion module,the introduction of attention mechanism and the introduction of receptive field enhancement module.The main research work of this paper is divided into the following three parts:(1)This paper first expounds the research on general target detection algorithms,and then classifies and summarizes the classic algorithms in the field of target detection according to the detection ideas and multi-scale target detection methods.In addition,in view of the difficulty of small target detection,this paper starts from the challenges faced by small target detection,and summarizes the solutions to improve the performance of small target detection.Finally,various datasets and evaluation criteria in the field of object detection are summarized.(2)In order to improve the detection performance of the SSD network model for small targets,this paper proposes a target detection algorithm based on feature fusion and attention mechanism.Shallow features have high resolution and contain more detailed information,but at the same time,their semantic information is poor and may contain more noise.However,after multiple convolutions,the high-level features have rich semantic information,but the resolution is low,and the perception of details is weak.In this paper,through the improved feature fusion module,three feature maps of different scales are fused to obtain intermediate layer features with richer semantic information.Additionally,this paper investigates the impact of different attention mechanisms on object detection performance.By introducing an attention mechanism,the network model pays more attention to the regions containing objects.At the same time,for the initially obtained intermediate feature layer,through the self-attention unit module,more useful information is captured for subsequent detection.(3)In order to improve the receptive field of small targets,this paper proposes a target detection algorithm based on receptive field enhancement.After the original image undergoes multiple convolution operations,the size of the receptive field corresponding to the pixels on the feature map of each layer is different.The larger the receptive field,the more global information the pixels in the feature layer contain.This paper uses the method of dilated convolution to obtain features with larger receptive fields and more information with a small amount of calculation.Experiments prove that the improved algorithm proposed in this paper based on the SSD network model makes more effective use of image information and obtains better features for detection.The detection accuracy of the two proposed detection algorithms on the PASCAL VOC dataset reached 80.6% and 81.0%,respectively,and the detection speed reached 41 frames per second.While satisfying real-time detection,it achieved higher detection accuracy.