Application Of The Random Forest Classification Algorithm In Edge Detection

As a classic image processing problem,edge detection is the basis of image segmentation,feature extraction,target recognition and tracking,and is widely concerned by researchers.Existing edge detection algorithms often need to adjust the parameters to deal with the transformation of the scenes,while it is difficultto achieve a good balance between the noise robustness and detection accuracy.Therefore,it is an important improvement direction of the edge detection algorithm to find a kind of algorithm with high adaptability and high detection precision.Based on the traditional edge detection algorithms,the paper analyzes the principle of traditional edge detection operator such as Roberts operator,Sobel operator,Canny operator and Laplace operator,and points out theirshortcomings.This paper expatiates the advantages of random forest training and classification,the realization process is simple and the model generalization ability is strong.Combining with the existing application of random forest in image processing,an image edge detection algorithm based on structured random forest is proposed:the image blocks and the corresponding real edge tags are used as the data sample set of the training random forest model;the information incremental theory and principal component analysis are used to find the optimal splittingfunction of each node in the forest;the integrated strategy is combined with multiple trees' classification results;finally,image blocks are classified to the corresponding edge tags by the trained random forest model.The experimental results show that the proposed algorithm based on structured random forestis highly adaptable,has excellent noise robustness and detection precision,and exhibits good performance in the application.The paper's characteristics and innovationsare summarized as follows:?This paper presents a concrete method to realize the structure of the input and output space of the random forest.The pixel values of the pixels of the original image block and the gray value of the pixels are extracted from the original image block.The edge tag is mapped by the mapping function to be directly calculated European distance binary vector.?The method of obtaining the optimal splitting function of each node in the random forest is as follows:the information gain is calculated before and after the introduction of the information increment.The entropy between the original sample set and the sample set after splitting is the information gain,The calculation of entropy depends on the classification of the sample set by one-dimensional principal component analysis,and the splittingfunction with the largest gain of information is chosen as the optimal splitting function.?A decision tree integration strategy is proposed for the structured random forest.The Euclidean distance between the binary vectors corresponding to the edge labels of each decision tree is calculated,and the core label is selected as the final output,instead of the voting method and the averaging method.?Combined with the oil bottle project,the edge continuity,detection rate and other edge detection quality evaluation factors are introduced in the paper.The subjective comparison and objective analysis are used to show that the algorithm is fast and effective.