Research On Object Detection Algorithm Based On Weakly Supervised Learning

Since the introduction of the convolutional neural network(CNN),object detection algorithm based on fully supervised learning has made great progress in the performance of network models,and has already been applied in the industry.Behind the powerful performance of object detection is a large amount of data annotation work,which will consume a lot of time cost.Therefore,object detection algorithm based on weakly supervised learning emerges as the times require.However,the object detection performance based on weakly supervised learning is low due to insufficient annotation information.In order to solve this problem,in the absence of instance-level annotation information,combined with related research work such as CNN,multi-instance learning(MIL),teacher-student distillation systems,etc.,an object detection algorithm based on weakly supervised learning(WSOD)is proposed,which uses only image-level labels and achieves fairly good results in the current field on key datasets.In order to further improve the performance of the algorithm,this paper adds a regression module on this basis.Aiming at the problems caused by MIL in the model training process,the loss function of the self-training network is improved,the proposal region screening algorithm in the WSOD problem is optimized,and a pseudo-label generation algorithm for regression operation is proposed.The algorithm finally achieved very good performance under the standard dataset.The main research contents and results of this paper are as follows:(1)Previous studies have proved that the WSOD problem can be classified as MIL problem,and the addition of MIL greatly improves the detection effect.From this perspective,this paper constructs an algorithm architecture consisting of a feature extraction network(FEN),a multi-instance detection network(MIDN)and a self-training network(STN).Next,based on this architecture,the feature extraction capabilities of several mainstream backbone networks are explored,and ablation experiments are used to find the most suitable network design mode for the WSOD problem.Then,combined with the application of attention mechanism in the field of computer vision,a common attention network model is added to the FEN,and an attention module suitable for the WSOD task is proposed based on the existing attention module.Finally,this paper improves the output of the STN to make full use of the supervision information at each stage,further improving the performance of the network.Comparative experiments on standard datasets show that the algorithm proposed in this paper has good performance in the WSOD task,which verifies the applicability of the innovation to this problem.(2)The lack of instance-level annotation information makes the WSOD algorithms unable to perform bounding box regression and inefficient,so this paper adds a regression module to the algorithm.The paper firstly introduces the different strategies of regression localization and its role in the traditional object detection problem.Then,the loss function used by the STN is improved to solve the problem of trapping in local optimum caused by the non-convex nature of MIL during the training process.Secondly,the scoring-based proposal region screening algorithm is optimized to improve the network’s detection of homogeneous multi-object images.Finally,a novel IOU-based pseudo-label generation algorithm is designed,focusing on solving the problem of instance-level label missing in the regression branch and improving the accuracy of pseudo-label generation.Experiments on key datasets show that the regression module plays an important role in the network,further improving the network performance,thus verifying the effectiveness and robustness of the algorithm in this paper.