Small Target Detection Method And System Based On Incremental Learnin

Object detection is widely used in our daily life,especially in the fields of autonomous driving,industrial inspection,and medical imaging.With the development of deep convolutional neural networks,many mainstream detectors have made remarkable progress on general data sets,but the small object detection task still faces many challenges.On the one hand,the number of pixels on the image is usually small due to the low resolution of small objects,which will cause the gradual loss of position information in the feature extraction process,and features that can be learned are very limited,so the detection algorithms exist some loopholes,such as low recognition rate,high leakage rate,and other problems.On the other hand,the current detection model is very dependent on the current data.It can only detect fixed types of objects and cannot incrementally adapt to new objects.Once the network is trained with a new dataset,new learning knowledge will overwrite old knowledge,and the model will forget how to detect the old objects(i.e.catastrophic forgetting occurs).In response to the above problems,the main research contents of this article are as follows:1.This paper proposed a small object detection method based on enhanced feature learning,and investigate two models based on single-stage networks.Meanwhile,a key point feature enhancement model and a hopping attention feature enhancement model were designed.The key point feature enhancement model uses deformable convolution to enhance the learning of small target features,which improves the fitting ability of the model and avoids the problem of losing small object location information caused by excessive downsampling.What’s more,it also improves the Gaussian function to make it generate the Gaussian distribution that can better reflect the true aspect ratio of the target,and improve the accuracy of the border regression.The feature enhancement model based on hopping attention is designed to strengthen the extraction of small object features.A new method was used to calculate the border loss and increases the penalty of the border loss for small objects to improve the localization of small objects.Results showed that this method can effectively improve the detection accuracy of small objects.2.In this paper,a small object detection algorithm based on label generation and selection was proposed.Using the old model detection results and the available new data labels,a label merging method based on border distance measurements was designed to make up for the missing old object information in the new data,which solved the problem that the small object detection model recognizes the old object as background and alleviated catastrophic forgetting.The attention residual module was designed to learn important features of small objects at different depths in the feature extraction stage,which further improves the accuracy of the model to detect old and new targets.The results indicated that the incremental detection method in this paper didn’t need to access the old object dataset,and it performs well in both new and old objects detection.3.This paper designed and implemented a small object detection system.This system designed multiple modules to meet the actual needs of users,which included four modules of user management,online detection,batch detection and data management.Based on the small object detection method proposed in this paper,an easy-to-use and well-interactive detection system has been realized.