Research And System Design Of Rice Image Analysis Algorithm Based On Deep Learning

Combining deep learning to achieve automatic detection of food grains is not only beneficial to reduce labor costs and duplication of efforts,but also to establish uniform standards and form a reliable database.Therefore,this thesis combines artificial intelligence with agriculture to study an automatic inspection system for the quality of rice grains,which can efficiently and accurately complete the segmentation of rice grains and perform quality inspection of batch samples based on statistical information such as grain shape,chalkiness and mold.This thesis focuses on the following work:Firstly,a rice image segmentation and classification dataset is constructed for the situation that there is no publicly available dataset of rice images in the academic community.A synthetic dataset generation method is proposed to support the automatic segmentation and classification tasks of rice images.For the quality and characteristics of rice image data,the dataset creation process is defined,and the annotation results of the segmented dataset are obtained by machine coarse annotation and manual fine annotation methods.In order to compensate for the shortcoming of insufficient data volume of datasets,a synthetic dataset generation method is proposed,which can simulate the rice distribution of real data and generate different types of datasets according to user requirements.At the same time,the corresponding segmentation annotation is automatically generated,reducing the manual annotation workload.The constructed rice image segmentation dataset contains 210 annotated images and 1400 synthetic images;the classification dataset is divided into 7 categories,including 1300 images.Secondly,to address the current situation that the effect of rice segmentation needs to be further improved,TISN is proposed and a synthetic dataset is used for training to achieve instance segmentation of rice grains.TISN uses ResNet+FPN for feature extraction.The transformer encoder with dynamic attention is used to learn the relationship between features,and the head prediction network is used for detection and segmentation,and a recurrent update mechanism is used to continuously update the prediction set.Experiments show that this network segmentation results in about 6%improvement over the baseline.Thirdly,a weakly supervised detection algorithm is proposed for the rice-grain classification task.While completing the classification,it realizes automatic annotation of the area of interest.The algorithm fuses MobileNet and class activation mapping algorithms to visualize and analyze classification results,and uses thresholds to determine the interesting regions of classification results,enabling weakly supervised detection of classification concern regions.Fourthly,a comprehensive analysis system of grain image is designed,which combines rice grain segmentation algorithm and rice image classification algorithm to effectively accomplish the task of rice quality inspection in the actual production process.