| Scene understanding provides the basis for the application of computer vision in many fields such as autonomous driving.Semantic segmentation is one of the key technologies in scene understanding.Semantic segmentation technology achieves deep understanding of the scene by predicting pixel-level classification results.With the rapid development of deep learning methods,convolutional neural networks have achieved competitive segmentation accuracy in semantic segmentation tasks.However,the high performance of most convolutional neural networks relies on the use of large-scale datasets to train the network,and in some professional fields such as military target analysis,it is difficult to obtain large-scale labeled datasets.Such problems limits the applicability of neural network methods.In response to the above problems,the few-shot learning method is proposed.The purpose of few-shot learning is to train a model with generalization ability using images of known categories,so that only a very small number of labeled samples can be used in the inference stage to achieve the understanding of unknown categories.However,the semantic segmentation accuracy of the existing few-shot learning methods still needs to be improved.In order to achieve a higher-precision semantic segmentation accuracy under the condition of few samples,this paper deeply studies the neural network semantic segmentation optimization technology based on few-shot learning.In this paper,the optimization algorithm of prototype in few-shot learning prototype network is firstly proposed.Then,in view of the problem that the existing methods have poor segmentation accuracy for small objects,the multi-scale feature fusion method under the condition of few samples is deeply studied.As RGB-D data are low-cost and easy to obtain,a few-shot learning multi-source feature fusion semantic segmentation network model using RGB-D data is proposed.The main research results of this paper include:(1)In view of the problem that the prototype in the prototype network does not fully utilize the sample information,this paper proposes a fewshot learning optimization prototype network model.The distance between nearest neighbors is calculated to assign different weights to the features in the prototype,and enhances the weight of the features that are conducive to the segmentation task to improve the performance of the network model.The mIoU of our model is improved by up to 8.2%compared with the baseline model PANet on the PASCAL-5i dataset,and is improved by 2.0%compared to other prototype network methods on COCO-20i dataset.(2)Aiming at the problem that the few-shot learning method has poor effect on small target segmentation,this paper proposes a few-shot learning multi-scale feature fusion network based on the spatial attention mechanism.The network extracts the rich spatial information from shallow features through the spatial attention mechanism.The spatial information is then fused with the deep features which rich in semantic information,and the obtained fused features can better guide the network to complete the segmentation task of objects of different scales.The network model can achieve the mIoU of 60.8%on the PASCAL-5i dataset,which provides a 3.3%improvement over the method REFs proposed in 2021.(3)Aiming at the problem of insufficient information due to the small number of samples in few-shot learning,this paper proposes to use RGBD data which is low-cost,easy-to-obtain and can share annotation data with color images to perform semantic segmentation tasks under the condition of few samples.Since the existing datasets contains insufficient scenes,a self-collected dataset HomeObjects-3i is set up.It contains 160 image data and 480 object mask annotations.The dataset simulates a semantic segmentation scenario for goods on a conveyor belt in a factory environment.After that,this paper studies the multi-source data fusion technology of few-shot learning.Through the visual analysis of color image features and depth image features,a feature fusion network based on the channel attention mechanism is proposed,and the feature quality is improved by enhancing the weight of the effective channels in the feature.The splicing method is used to complete the feature fusion to preserve the complementary information in different features.Compared with the baseline model PANet,the network model can achieve up to 7.7%mIoU improvement on the public dataset Cityscapes-3i,and achieves 66.9%mIoU on the self-collected dataset. |
