Font Size: a A A

Research On Plant Recognition Algorithm Based On Feature Fusion Network

Posted on:2021-05-31Degree:MasterType:Thesis
Country:ChinaCandidate:C Y NieFull Text:PDF
GTID:2393330620473112Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
Plant identification has important applications and value in the fields of forestry conservation management,forest resource research and natural environment monitoring.In order to solve the problem of unreliable single organ recognition and unbalanced distribution of plants in nature,we propose a plant recognition method based on feature fusion network.We took the plant images of Plant CLEF2016 as the research object,and studied the feature fusion network,loss function design and network feature enhancement based on salient features.The main research contents and conclusions of this paper are as follows:(1)Feature fusion network based on residual network.The Res Net50 network is used as the basic feature extraction network.The convolution layers with the same size output features in Res Net50 are classified into the same stage.And a multi-layer feature fusion module and a multi-scale feature fusion module are designed.In the multi-layer feature fusion module,the high-level features and low-level features of the same stage was fused through short connection operations to obtain fusion features;In the multi-scale feature fusion module,the fusion features of each stage were used as supervisory information to recalibrate the low-level features of this stage.The calibrated features in each stage are respectively subjected to global average pooling and concatenating operations to obtain feature vectors of plant images.The experimental results show the superiority of the feature fusion network.(2)Feature fusion network based on improved loss function.Based on the plant recognition task in this paper,a loss function is designed to replace the cross-entropy loss function in the feature fusion network.The additive intervals to compress the cosine distance was used in the improved loss function,so that the classification interval between different categories is increased and the features within the cluster are gathered to alleviate the problem of small differences between the sample classes.In addition,by adjusting the weights of different categories in the loss function,the network is more fully trained on the small sample categories,and the problem of sample imbalance is alleviated.The effectiveness of this method was verified on Plant CLEF2016,a plant data set with extremely uneven sample distribution.(3)Network feature enhancement based on salient features.After the feature fusion network extracts the plant image features,it is enhanced by saliency features.First,the plant image features are extracted through the multi-feature fusion network,and the salient feature extraction of the plant image features is performed using the K-means-based salient feature extraction algorithm to obtain the salient features of the image.Then the plant image features and salient features was fused by adopting adaptive weighted feature fusion strategy.Finally,the fusion features are classified and recognized.In summary,based on feature fusion networks and the optimization of feature fusion networks,this paper realizes plant identification under the condition of multiple organs and imbalanced sample distribution.The experimental results have verified the effectiveness of the proposed method,indicating that the research results of this paper have certain practical significance,and have certain practical significance for plant protection,intelligent agriculture,etc.
Keywords/Search Tags:Plant recognition, Feature fusion, Improved loss function, Significance feature, Sample imbalance
PDF Full Text Request
Related items