Research On Farmland Crop Recognition Method Based On Deep Lab V3+ And Superpixel Segmentation

With the rapid development of technology,the collection of agricultural situation images has become more and more convenient.Accurately and efficiently processing a large number of agricultural situation images is the key to achieving accurate agricultural situation monitoring.In order to solve the problem of crop recognition for VGI farmland images in complex scenes,this paper proposes a farmland crop recognition method that combines DTEMA semantic segmentation and superpixel segmentation,which is based on the Deep Lab V3+ semantic segmentation framework.The research includes the following innovations:(1)This paper proposes a multi-scale texture feature enhancement module based on Gabor filter.DTEMA uses Deep Lab V3+ as the basic structure.Since the color features of different crops are very close,a multi-scale texture feature enhancement module is introduced into the model’s encoder to increase the proportion of texture features in the feature map to improve the accuracy of crop species recognition.(2)This paper proposes a multi-layer attention fusion module.The module separately obtains the attention features of different levels of feature maps in the backbone,which strengthens the dependences of the global feature points and the feature channels,thereby improves the average accuracy of the model in identifying crops.(3)This paper proposes a fusion algorithm based on threshold voting.In order to solve the disadvantage of edge information loss in the semantic segmentation model of encoding and decoding structure,this paper uses the threshold voting algorithm to merge the results of the DTEMA semantic segmentation with the SLIC superpixel segmentation results,which further improves the accuracy of the model’s recognition of crop edges.The experimental results show that the introduction of texture feature enhancement module optimizes the performance of the model and improves the accuracy of the model in crop classification.The multi-layer attention fusion module is introduced to improve the performance of the model,and the model has significantly improved the recognition effect of farmland images with scattered plots.Using the algorithm based on threshold voting to merge semantic segmentation and superpixel segmentation improves the segmentation effect of crop edges.