Small Object Detection Algorithm Based On Convolution Neural Network

With the continuous development of convolutional neural networks,object detection has become an important application field of convolutional neural networks.Many current excellent algorithms have high average accuracy on general data sets.The accuracy of small object detection of the CornerNet algorithm is higher than that of other algorithms in the same period.However,compared with large and medium objects,the accuracy of small object detection is relatively low.Because small objects have problems such as too small size,many noises,unobvious features,and blurred edge information,the small objects merge with the background,which is easy to cause misdetection and missed detection,and the detection effect is poor.By consulting related literature,the current object detection algorithm has three main reasons for the low accuracy of small object detection: insufficient feature map fusion leads to weak semantic information;lack of multi-scale feature maps for subsequent network detection;setting more dense anchor frames for sampling.Since the CornerNet algorithm transforms the anchor frame problem into a key point problem,in order to solve the problem of low accuracy of small object detection in the CornerNet algorithm,this paper proposes a small object detection algorithm that combines the stepped fusion feature pyramid and CornerNet.The network is improved and an improved feature pyramid is added to improve the accuracy of the algorithm for small object detection.First of all,this article improves the backbone network(Hourglass-104)of the CornerNet algorithm.Hourglass-104 is composed of multiple residual blocks.Each residual block is composed of two convolution kernels with a size of 3*3.Reduce the number of parameters and shorten the reasoning time.Under the premise of ensuring accuracy,the residual block is changed to two convolution kernels with a size of 1*1 and a convolution kernel with a size of 3*3 in between.To speed up the reasoning time of the backbone network.Then,the method of combining the feature pyramid and the CornerNet algorithm is applied to small object detection.In order to test the influence of the number of feature pyramid layers on the accuracy of small object detection,the four feature maps after the fusion of the second half of the backbone network are extracted.A4-layer feature pyramid is formed,and the smaller-sized feature map in the 4-layer feature pyramid is convolved twice to obtain two convolutional feature maps to form a6-layer feature pyramid;finally,the 4-layer feature pyramid and the 6-layer feature pyramid are combined On the basis of the upper and lower fusion,the feature pyramid performs inter-level fusion and bypass connection,so that the fused feature map has rich semantic information,so as to efficiently complete small object detection.The improved algorithm is compared with the current mainstream algorithms CornerNet,Faster R-CNN,and Retina Net on the same MS COCO data set.The results show that when the improved algorithm detects small objects,the accuracy of small object detection is greatly improved.The inter-level fusion feature pyramid can effectively fuse high-level and low-level feature maps on the CornerNet algorithm,so that the fused feature maps have strong semantic information,and improve the accuracy of the CornerNet algorithm for small object detection.