Research On Small Sample Target Transfer Detection Based On Contrastive Learnin

Object detection(OD)algorithms have developed by leaps and bounds with the emergence of deep convolutional neural networks.However,due to the difficulty of labeling large-scale training sets,the promotion of OD applications needs to solve the problem of few-shot learning,that is,after pre-training of base-class samples,use a very small number of new-class samples for simple model migration to efficiently detect new-class instances.Recently,some researchers have extended contrastive learning,which is outstanding in the field of semi-supervised learning,to the field of small-sample target detection.However,due to the memory requirements of the object detection framework and the limitations of the anchor frame mechanism,contrastive learning has not been fully utilized.Furthermore,the transfer of detection frameworks from base classes to new ones is prone to catastrophic forgetting and conflicts between regression and classification tasks.To this end,this paper studies the adaptive improvement of the small sample object detection framework based on contrastive learning,and proposes two effective improved algorithms.First,a few-shot object detection method based on low-resource contrastive learning and online Fisher matrix update is proposed.Aiming at the catastrophic forgetting problem when the model migrates to a new class,the elastic weight curing(EWC)is extended to the field of few-shot object detection by using the MSE statistical evaluation,and by designing an online soft constraint,the Fisher information is used in the training Evaluate parameters within a sample batch to constrain or encourage transfer model parameter updates.Furthermore,a momentum update-based inter-batch storage mechanism is employed to alleviate the memory strain caused by the contrastive learning module when driving a large number of contrastive encodings.Finally,a lightweight heterogeneous decoupling head is designed to de-entangle information between regression and classification tasks.The method was verified experimentally on the PASCAL VOC and COCO datasets.Compared with the original contrastive learning baseline for few-shot object detection,the average detection performance of the new-class improved by 1.3%～15.1%,while the average detection performance of the base-class improved by up to 7%.reach a competitive level.Second,a new few-shot object detection method based on multivariate prototype clustering contrastive learning and task decoupling is proposed.Utilizing the highefficiency coding ability of contrastive learning in the few-shot object detection scenario,a multi-prototype feature learning mechanism based on contrast update is further designed to reduce the memory shortage and sampling bias of base-class samples during contrastive learning coding.At the same time,when the model is transferred to a new-class,the shared network weights of the regression and classification tasks inevitably cause information entanglement,resulting in the inability to accurately model the two tasks at the same time.To this end,this paper proposes a task-decoupled head network architecture with feed-forward separation and feedback gradient decoupling to enhance discriminative feature learning for a single task.Similarly,this method is also experimentally demonstrated on the PASCAL VOC and COCO data sets.Compared with the existing few-shot object detection baseline based on contrastive learning,the average detection accuracy of the new-class of this method is increased by 3.6% to 16.1%,which also reaches a competitive level.