Research On Plant Depth Maps Recovery Based On Target Features

Depth data is of vital significance for 3D reconstruction,the appearance of depth camera makes that it is easier for users to acquire depth data and makes 3D reconstruction more simple and fast.However,depth map acquires by depth camera often appears missing and wrong pixels,especially the plant depth maps,the depth data often lost in the position of leaves and branches.And it's not ideal to use the common depth maps recovery methods to repair the plant depth maps.Considering the depth maps of plants provided by depth camera are incomplete,and common filtering methods can't repair the plant depth maps accurately,we propose a plant depth maps repair method which is based on target features.This method is divided into plant depth map recovery based on spatial fitting,plant depth map recovery based on support vector machine and spatial transformation.The main contents of this paper are as follows:1.Pretreatment of plant image.First,we register plant maps,and use the color image segmentation algorithm which based on color and spatial information to segment the plant image.We extract the information of each target region's internal pixel points as sampling points and missing points,and then based on the number of sampling points and missing points to determine which recovery method is suitable for this region.2.Using spatial fitting to repair the plant depth map.We use moving least-square method to fit each region's spatial equation,and then put the wrong and missing depth pixels into this equation,calculate pixel's depth value to repair small area's wrong and missing depth pixels in the regions.3.Leaf image data set generation and feature extraction.Because leaf map data sets are incomplete,we increase the number of data sets.In addition,we extract features from leaf map to predict the leaf type and retrieve the best matching leaf image.4.Using Support Vector Machine and spatial transformation to repair the plant depth map.Due to the spatial fitting method cannot accurately fit the region's equation which has large area missing depth pixels,we use SVM to train the leaf image data set,and then predict the classification of plant leaves,retrieve the best matching leaf by similarity measurement.According to the spatial information of the two leaves to calculate the accurate depth values to repair large area's wrong and missing depth pixels in the regions.Experimental results show that the proposed method can accurately repair the wrong and missing depth data in the plant's depth map.It achieves better performance for plant depth maps recovery,and protects targets' edge information efficiently.