Research On Deep Target Detection Algorithm Based On Semi-supervised Learning And Active Learnin

Object detection is a primary task in the field of computer vision,which plays a vital role in higher-level tasks such as instance segmentation and object tracking.Due to the rise of deep learning,the performance of object detection has been greatly improved,but obtaining a highperformance detection model often requires a large amount of labeled data and computational costs.Nowadays labeled data is scarce and expensive,while unlabeled data is cheap but rarely used.Therefore,how to use unlabeled data to improve the performance of detection models has become a current research hotspot.The topic of this paper is to make full use of unlabeled data to improve the performance of the detection model when using a small amount of labeled data and further reduce the computational energy consumption of the model on this basis.To make full use of unlabeled data to alleviate the problem of insufficient labeled data,researchers have been exploring methods using semi-supervised learning or active learning.Among them,active learning selects valuable unlabeled data for labeling,while semisupervised learning utilizes information in both labeled and unlabeled data.This paper considers the complementarity between active learning and semi-supervised learning and uses these two methods in combination to apply to object detection algorithms that uses a small amount of labeled data and a large amount of unlabeled data to further improve the performance of the algorithm.Semi-supervised learning uses data enhancement and consistency criteria,combining labeled and unlabeled data to train detection models.Active learning combines informativeness and representativeness to query samples that can best improve model performance on unlabeled data,and then mark them again.Train to improve the performance of the model.Experiments fully prove the effectiveness of the method proposed in this paper and use it as a basic framework for follow-up research.Aiming at the problem of reducing the energy consumption of the model,this paper further explores the possibility of using spiking neural networks to realize energy-efficient active semisupervised object detection algorithms.Compared with the traditional neural network that requires continuous calculation and storage,the spiking neural network is sequential,and pulses are emitted only when the input signal changes,which greatly reduces the energy consumption of calculation and storage.This paper uses the method of converting the deep neural network into a spiking neural network to apply the spiking neural network to the target detection task and constructs a target detection model that combines semi-supervised and active learning.The experimental results show that the method achieves better detection performance with only a small amount of labeled data,which saves the cost of labeling and greatly reduces the energy consumption of the model.