Research On Visual Counting Method Based On Deep Learning

Object counting is a common task in daily life,and it is also a basic problem in many scientific fields.With the maturity of imaging technology,image and video data has been growing rapidly.How to obtain the number of objects quickly and accurately is a crucial practical issue.Driven by deep learning,the use of computer vision technology has improved the speed and accuracy of object counting.However,the application scenarios of object counting are extensive and complex.This work studies the visual counting method based on deep learning for single-frame images and multi-frame images,and applies research in agriculture and marine fields.As one of the important food crops,wheat production is closely related to people's lives.With the help of computer vision technology,it is a new attempt to estimate wheat yield by analyzing images of wheat.In this work we collect images of wheat at the filling stage,and count the number of wheat ears per unit area through the object detection and counting method based on deep learning.With counting the number of wheatears,we accurately complete the task of wheat yield estimation.The taxonomic composition and abundance of plankton in the ocean observation.Plankton observers need to build a plankton image database and look for automated plankton identification statistical algorithms.The dark-field flow-imaging system for plankton can quickly acquire a large number of images,which is suitable for the observation of plankton in large areas of sea.This paper studies the image enhancement method based on the multi-frame image super-resolution algorithm and designs a plankton tracking and counting method based on the object association method.In view of the above questions,this work is summarized as follows:1)Wheat yield estimation method based on single frame image object countingA method for estimating wheat yield based on the object counting in the single-frame images is proposed.First,we establish a wheat ear detection dataset.Each image in the dataset contains more than 100 wheat ears.Then we use the YOLO v3 to detect and count the wheat ears in these images.Finally,the wheat yield is estimated according to the calculation formula of wheat yield.The experimental results on our test dataset prove that our wheat ear counting and yield estimation based on single-frame image is feasible.2)Plankton image enhancement method based on multi-frame super-resolutionAn image enhancement method based on multi-frame image super-resolution is proposed.Our method first generate background images and remove the influence of impurities in the background on image analysis.Then we adopt adaptive histogram equalization that limits the contrast(CLAHE)to realize the further optimization of the image.Finally,the super-resolution method for image sequence(EDVR)based on deep learning is used to improve the resolution of the image.This method improves the visibility of the image and provides high-quality images for the plankton image database.3)Plankton detection and counting method for multi-frame imagesA plankton detection and tracking method for Continuous Plankton Images from Dark-field Flow-imaging(DFCPIs)is proposed,and the classification and counting of plankton in the image are completed.Firstly,the images are preprocessed.Then we use the modified Faster R-CNN to obtain regions of interest(ROIs)on all images,extracting the category and location of plankton.Finally,we achieve the matching of the same individual in the video by object association.Experiments show that this method has a good tracking and counting performance.