Research Of Semantic Segmentation Algorithm Based On Convolution Neural Network And Its Application In Image Recognition Of Sweet Pepper

Sweet pepper is widely planted in greenhouse,which has rich nutritional value and economic value.Picking is an important link in the production process of sweet pepper.As the development trend of modern agriculture,automatic picking has been widely concerned.With the rapid development of artificial intelligence,the research of greenhouse sweet pepper picking robot has become a hot spot.Machine vision system is an important part of picking robot.Among them,image recognition technology is particularly important for accurate recognition of the target.However,the traditional image recognition methods rely on the application of artificial feature extraction and laboratory background,so it is difficult to achieve high image recognition accuracy in the unstructured environment of greenhouse.At present,the deep learning method represented by convolution neural network has been successfully applied in the field of agricultural image recognition.However,the lack of relevant annotation data sets and effective depth recognition model limits the development of deep learning in the field of agricultural computer vision.In addition,in order to improve the picking efficiency and ensure the freshness of fruits and vegetables,it is necessary to pick them during the day and at night.Therefore,this paper takes the sweet pepper image of greenhouse at night as the research object,studies the semantic segmentation algorithm based on convolution neural network and transfer learning,and applies it to the sweet pepper image recognition task.The specific contents and conclusions are as follows:(1)Taking sweet pepper synthetic image as the research object,the semantic segmentation algorithm was studied on the premise of sufficient training data.The existing convolution neural network model has some problems,such as low segmentation accuracy,poor robustness and unclear boundary segmentation.To solve these problems,a semantic segmentation model suitable for sweet pepper segmentation was selected as the benchmark network through experiments.Based on the fullyresolution residual network,the dilated convolution was integrated to increase the receptive field of the network,and the position attention module was combined to capture the global dependence in the spatial dimension.The results showed that the proposed model FRRN-DCov-PAM could achieve accurate segmentation and recognition of sweet pepper synthetic image at night.Compared with the original model and other semantic segmentation models,FRRN-DCov-PAM had higher recognition and segmentation accuracy,and its mean PA and mean IOU on the test set could reach97.94% and 78.88%,respectively.(2)Taking sweet pepper empirical image as the research object,the transfer learning strategy and the adversarial learning algorithm were studied in the case of insufficient training data.Aiming at the problem that there were few empirical images and effective network models could not be trained,a transfer learning method based on fine-tuning pre-trained model was proposed to transfer parameters of the optimal model in the research content(1).Due to the over-fitting phenomenon in the actual training process,Cycle-GAN was used to convert the synthetic image to obtain a training set closer to the empirical image.Then,the proposed domain adaptation network based on CNN was used to transfer the output prediction distribution of the source domain to the target domain,and the converted synthetic images were used for the training of the network,which solved the domain adaptive problem of semantic segmentation for sweet pepper image.The results showed that the proposed domain adaptive segmentation algorithm could achieve accurate segmentation and recognition of sweet pepper empirical image,and the mean Io U of the optimal model on the test set reached62.38%.In conclusion,the improved model FRRN-DCov-PAM and domain adaptive segmentation method proposed in this paper can achieve accurate segmentation and recognition of sweet pepper in greenhouse at night.This research can provide theoretical support for improving the visual performance of sweet pepper picking robot,and provide reference for solving other image recognition problems in the agricultural field,which has certain theoretical research significance and application value.