Research On Terrain Classification Method Of Photovoltaic Power Plant Based On Optimized Visual Bag Of Words Model

In recent years,photovoltaic energy in new energy has been vigorously developed in China,and the operation of a large number of photovoltaic power stations has brought many problems.The long-term exposure of photovoltaic modules to the outside seriously affects their power generation efficiency and service life due to dust accumulation on the surface.It is crucial to use autonomous mobile cleaning robots to timely clean the dust accumulation on the module surface.The road surface of photovoltaic power plants has non geometric hazards,including soft sand,muddy land,smooth grassland,uneven gravel,and uneven terrain.Therefore,mobile cleaning robots need to recognize and classify the terrain environment when working independently in photovoltaic powe r plants,in order to adopt corresponding walking strategies to ensure that the robot safely crosses the terrain.This article focuses on the recognition and classification of different types of terrain in photovoltaic power plants.Based on machine vision,methods such as terrain image acquisition,feature extraction,and classification model construction are studied.The main research content is as follows:(1)Analyzing complex terrain scenes of photovoltaic power plants,Under multiple conditions,five common types of photovoltaic power plant terrain image samples were collected using cameras.Firstly,perform image preprocessing such as denoising and resolution reduction on terrain images,and construct the Terrain-5 dataset.Then,based on the color and texture features of different terrain images,HSV color features and rotation invariant LBP texture features are used to describe the color and texture of the terrain.(2)In view of the shortcomings of the traditional visual bag-of-words model model,which ignores the spatial position information of visual features in terrain classification and has weak feature discrimination,the spatial pyramid model is introduced to improve the spatial pyramid structure according to the feature dimension of the image extracted from the model structure is too high and the characteristics of the terrain prominent area.Optimize the traditional spatial pyramid model with a scale of l=3 for region division,subdivide the intermediate regions,make the extracted features r epresentative,and reduce the dimensionality of spatial pyramid features.The experimental results show that the terrain classification effect based on the optimized spatial pyramid visual bag-of-words model model is better than the spatial pyramid visual bag-of-words model model and the traditional visual bag-of-words model model.(3)Aiming at the problem that spatial pyramid visual bag-of-words model model ignores color information and has no obvious texture expression when classifying terrain,a terrain classification method based on DCA feature fusion of spatial pyramid model is proposed.The DCA feature fusion framework is designed based on the DCA algorithm.First,three groups of transformation features are constructed by the DCA algorithm for the optimized spatial pyramid visual bag-of-words model model histogram features,HSV color features and rotation invariant LBP texture features,and then the three groups of transformation features are fused by the concatenation method.The final experimental results show that the fused features have low dimensionality and strong discrimination ability,and have higher accuracy in terrain classification than single features.(4)Aiming at the problem that a single classification model is prone to fall into overfitting and low generalization ability,a terrain classification method based on Stacking integrated multi classification model is proposed.SVM,KNN,and RF were selected as the base models for the integrated model,while Light GBM was used as the metamode l for the integrated model.Three different two-layer integrated classification models were constructed using the Stacking combination strategy,with fused features as input for the integrated model.Through comparative analysis of terrain classification e xperiments,the Stacking integrated classification model performed better and performed more robustly in terrain classification.