Research On Image Recognition Methods For Close Range Large Viewing Area

The container unloading truck is a device that enables quick unloading of cardboard goods from containers.The truck is manually controlled to locate the center of the stacked cardboard boxes and then to grab them.However,this manual positioning method is difficult to achieve for densely packed and numerous stacks of cardboard boxes and is prone to errors.Furthermore,conventional machine vision methods are unable to obtain panoramic images of large areas of stacked cardboard boxes.Therefore,this thesis proposes a recognition system based on image stitching and deep learning to achieve rapid and accurate recognition of stacked cardboard boxes in large areas,thereby reducing positioning errors and improving unloading efficiency.The main research content of this thesis is as follows:(1)The working conditions,main structure,and unloading process of the container unloading truck are analyzed to determine the requirements for the recognition system.Multiple cameras are used in parallel to acquire images of the stacked cardboard boxes,and image stitching and deep learning techniques are used to obtain panoramic images of the cardboard stacks and to recognize the cardboard boxes.(2)In image registration,traditional image registration methods extract feature points from the entire image area,resulting in too many interfering feature points and extended detection time.This thesis introduces the Hu invariant moment to find the overlapping area between images.Experiments show that extracting feature points only from this area can reduce the number of interfering feature points,shorten detection time,and maintain the accuracy of the original algorithm.Then,the SURF algorithm is used to extract feature points,and the Kd-tree algorithm is used for coarse feature point matching,followed by fine matching using the RANSAC algorithm.After matching,the perspective transformation matrix is calculated based on the fine matching results,and the images of the cardboard stacks are transformed to the reference image coordinate system using the perspective transformation matrix.(3)In the fused image after registration,the global differential of the seam line and the improved best seam line algorithm are used to obtain a more accurate seam line than the original algorithm.Then,the multi-resolution image fusion algorithm is combined to avoid the stitching gaps,missing image information,and ghosting problems caused by traditional fusion algorithms.(4)For the recognition of stacked cardboard boxes,initial datasets are established using images of container and warehouse cardboard stacks,and the dataset is expanded using data augmentation techniques.The YOLO v5 algorithm is used as the basic recognition network,and the loss function and backbone network are improved.By comparing the training results of the network models,it is found that the improved algorithm is superior to the original algorithm in terms of precision,recall,and average precision.The improved algorithm is used to recognize the images obtained by different fusion algorithms,and it is found that the recognition effect of the improved fusion algorithm is better.(5)Through simulation experiments,the effects of different numbers of lenses,different brightness environments,and different IOU thresholds on the recognition system’s operating time and recognition accuracy are explored,and the experimental results are analyzed.It is found that the recognition system can meet the design requirements in terms of operating time and recognition accuracy when four lenses are used under brightness environments between10 lux and 150 lux and IOU thresholds between 0.5 and 0.9.