Purple Soil Image Segmentation Based On Custom Measure

The high-quality segmentation of the purple soil visual image collected in the wild natural environment on the research target—the purple soil area image—can effectively avoid the influence of the background on the machine vision to identify the purple soil type.Therefore,the primary work is segmentation and extraction of the target purple soil region in the purple soil image.This thesis analyzed the distribution characteristics of foreground pixel values in the purple soil image in RGB space.A custom measure was constructed by selecting a feature.The initial segmentation image is obtained based on a custom measure and a density peak optimization model.The final segmentation and extraction of the purple soil image is realized by eliminating the discrete areas of the background and filling the empty areas of the foreground by post-processing algorithms.The combination of the improved custom measure and the gray-level connected density clustering algorithm make segmentation and extraction results of the purple soil image better than those before the improved algorithm.The main work of this thesis was as follows:(1)This thesis firstly used the distribution characteristics of pixel values in the RGB color space to calculate the conditional probability of multidimensional random variables.The optimal model was used to extract the boundary of the color feature value,and the segmentation function was used to customize the segmentation measure.Then,the custom segmentation measure was used to establish an optimization model based on the density peak idea.The local segmentation threshold was obtained by optimizing the model with two iterations to minimize the difference,and the intra-class variance minimization model is used to obtain the optimal segmentation threshold from multiple local segmentation thresholds to achieve the initial segmentation of the purple soil image.Finally,the recursive algorithm of pixel four-neighborhood connected labeling was used to mark the connected areas of the purple soil region image containing holes to eliminate the background discrete areas in the initial segmentation image.In the same way,marking the connected areas of the background to fill the holes in the purple soil area image realized the extraction of purple soil areas in the image.Finally,the feasibility of complete segmentation of the soil region in the purple soil image was verified by experiments.(2)We did not consider the uniqueness of the mapping relationship for the above-mentioned custom measures.This thesis combines piecewise linear mapping relationship and neighborhood cumulative probability mean processing to improve the custom measure.The multiples of maximum in the three-channel means was selected as the total mapping interval,and the three-channel means and the sum of the averages were used as the interval nodes.The pixel values of the purple soil color image were mapped one-to-one under the maximum boundary conditions.And the mean processing with the neighborhood cumulative probability make the improved segmentation measure maintain a certain similarity.For the density peak optimization model,the spatial information in pixels was not considered.In this thesis,gray value and its dynamic boundary were used to improve the gray connectivity density,and a maximum clustering optimization model was established to iteratively select the optimal clustering center to realize the initial segmentation of purple soil images based on gray connectivity density clustering.In the same way,Eliminating the discrete areas of the background and filling the holes in the purple soil region image achieve the final segmentation of the purple soil image.By calculating the comprehensive evaluation index F1,the Adjusted Rand Index(ARI)and the Normalized Mutual Information(NMI)to evaluate the segmentation extraction results,the results show that the improved segmentation method and average segmentation accuracy are better than the segmentation methods before the improvement.