Purple Soil Image Classification Based On Multi-Feature Fusion

The classification of purple soil image based on computer vision can help agricultural workers working in the field to identify the type of soil scientifically,which has a positive guiding effect on agricultural production.In order to improve the accuracy of purple soil image classification,the paper studies purple soil image classification based on multi-feature fusion.The main work of the paper includes the following three aspects.Firstly,extract the color features and texture features based on artificial design to classify the purple soil images.In order to reduce the influence of uneven brightness of natural light and shade on the color features of the image,the color information of the purple soil image with shadow in the RGB color space and HSV color space is analyzed,and it is found that H component color information in the HSV color space is less affected by uneven brightness of natural light and shade.Extract H component histogram features,H component color moments,and simultaneously extract GLCM features,LBP features,and RGB histogram features of purple soil images,and classify purple soil images using SVM classifiers.The simulation experiment results show that the classification accuracy using H component histogram features is the highest,which is more distinguishable than the other four groups of features.Secondly,optimize ResNet50 model to classify purple soil images.By comparing and analyzing the performance of AlexNet,VGG16,Inception-v3 and ResNet50 models for purple soil image classification,the simulation results show that ResNet50 model has the highest accuracy and the less training time,which proves the advantages of ResNet50 model for purple soil image classification.In response to the problem that the purple soil image data set is small and the accuracy of classification using ResNet50 model is not high,a transfer learning method is introduced.At the same time,a 128-dimensional fully connected layer is added after the global average pooling layer of ResNet50 model,and Dropout technology is introduced,and the ReLU-Mish function is designed to replace the ReLU activation function.Simulation experiments show that the classification accuracy using the optimized ResNet50 model has improved,and the learned features are more distinguishable.Thirdly,combine multiple features to classify purple soil images and apply.In order to describe the features of purple soil images more fully,the H component histogram features and the deep features extracted by optimized ResNet50 model are combined,and the combined features is 156 dimensions.To solve the problem that there may be redundant features between the combined features,and in the small sample purple soil image classification task,the feature dimension is too large and easy to overfit,the Relief F algorithm is used to calculate the weight of each feature in the combined features.The feature with a large weight has a positive effect on the classification performance.According to the cumulative contribution rate of feature weights,fusion features are selected to classify purple soil images.Simulation experiments show that the accuracy of purple soil image classification based on fusion features is higher than the classification accuracy based on H component histogram features,deep features extracted by optimized ResNet50 model,and combined features.It proves the effectiveness of multi-feature fusion for purple soil image classification.Then,the purple soil image classification model based on multi-feature fusion is applied to design and implement the purple soil image classification system.