Research On Tiny Object Detection Algorithm Based On Deep Learning

Object detection technology is an important branch of computer vision.It has been widely used in medical diagnosis,unmanned driving,remote sensing image recognition and other fields.With the development of deep learning,the detection algorithm has achieved good results in large and medium objects.Because the tiny has the characteristics of low spatial resolution,small coverage area and few available feature points,the tiny object detection algorithm based on deep learning still has the problems of missed detection,false detection and low positioning accuracy.Based on the above problems,a tiny object detection network named Large Kernel YOLO(LKYOLO)is proposed.The specific research contents are as follows:1.Large Kernel(LK)module is designed to solve the problem that tiny features are difficult to extract,furthermore,LK module makes use of the feature of large convolution kernel to cover a wide range of feature information,which strengthens the range of receptive field obtained by the network in shallow layer,thus increasing the long-range modeling ability between channels.The experimental results show that the large convolution kernel can effectively enhance the range of receptive field and has stronger inductive bias ability,which proves that the LK module effectively extracts the features of tiny objects and further improves the recognition accuracy of tiny objects.2.In order to solve the problem that tiny objects cannot be well fused on the feature map,a Large Kernel Spatial Pyramid Pooling(LKSPP)structure is proposed,which adopts a multibranch structure and adds a LK module to the branch.Multi-branch and large convolution kernels are used to capture the nonlinear features in the image,thereby enhancing the expression ability of the network and reducing the possibility of overfitting in the network,so that the network can better fuse the details.Experiments show that LKSPP effectively integrates tiny objects features and improves detection accuracy.3.In order to solve the problems of difficult accurate positioning of tiny target,low accuracy of anchor frame matching,and slow convergence of training,the Segment IoU(Segm IoU)loss function on the basis of Efficient IoU(EIoU)is proposed in this paper,and the Segm IoU regression is integrated into the positioning loss function.In each stage of Segm IoU,there is only one variable to control the position relationship between the two boxes,so as to achieve the effect of rapid convergence of the model.Experiments show that Segm IoU converges faster than EIoU.4.In order to solve the problem of missing features in the existing lightweight networks,the structure heavy parameterization method of Rep VGG is applied to the deep separable convolution model in this paper,and the multi-branch structure in the inference network is transformed into a single structure by the principle of matrix operation.In this way,the network training model and the inference prediction model are separated to accelerate the inference speed.Experiments show that the structural re-parameterization method accelerates the speed of the model in inference and prediction.Based on AI-TOD remote sensing image and Tiny-person maritime rescue image dataset,LKYOLO network and structural reparameterization are validated in this paper.The experimental results show that LKYOLO is superior to other target detection networks in detection accuracy and detection speed and has strong practicability in tiny object detection task.