Research And Implementation Of Small Object Detection System Based On Data Enhancement And Scale Normalization

With the development of deep learning neural network,object detection algorithm has become a popular deep learning direction.Object detection is the process of determining where objects are located in an image,and which class each object belongs to.Although the industry has made great progress in object detection tasks,there are still many challenges in small object detection:(1)In the data set of object detection,small object objects often only account for a small part,the number is not enough,and the distribution is uneven,resulting in insufficient learning of small objects and poor accuracy;(2)small objects have low resolution on the original image,and in feature weak in representation,it is difficult to obtain effective features similar to large objects;(3)There is a lack of a system platform that specializes in small object detection and can select two different small object detection algorithms according to actual scenarios.In order to solve the above problems,this paper proposes the following points:(1)First,this paper proposes a dynamic adaptive data-enhanced small object detection algorithm.The adaptive method can enable the small object detection network to adapt to the training data set.The input image dynamically adjusts the method and magnitude of data enhancement,making the object detection network more focused on small object detection and improving the detection performance of the network.We also propose a two-stage small obkect detection model that introduces a multi-head attention mechanism and anchor-free frame to improve the difficulty of expressing finegrained features of small objects and the inaccurate frame of region proposals.This algorithm is suitable for the case where the data set does not have too many small object features.Compared to SOTA,the performance is improved by 4.8%.(2)Sencond,this paper proposes a contextual multi-scale normalization small object detection algorithm based on adversarial learning,which performs generative adversarial learning on the low-level feature maps of the feature pyramid,uses the generator to generate small object super-resolution feature maps,and combines them with the real feature maps of large objects.It is sent to the discriminator for discrimination;while the high-level feature map introduces a balance module,using the features of different levels to make the receptive field pay attention to the non-local information and context information of the target object.The low-level feature map information is combined with the high-level feature map to generate super-resolution feature information of small objects to improve the performance of small object detection tasks.Compared to the baseline Faster RCNN,the performance is improved by 8%.(3)Finally,based on the above two algorithms,this paper designs and builds a small object detection system based on data enhancement and multi-scale normalization,deploys the above self-developed small object detection algorithm,and designs an interactive interface.Users can upload images to complete different small object detection tasks.This system has based on the major project"Advanced Machine Learning Theory Research for Open Environments" of the "New Generation Artificial Intelligence"project in Science and Technology Innovation 2030,and has practical significance in the open environment scenario of Hebei State Grid.