| Machine vision-based target recognition and counting can effectively solve the drawbacks of lengthy labor-intensive and low recognition rate in manual identification and counting,which is a crucial step to fully realize the intelligentization and automation of agricultural machinery.Most of the work in the agricultural field needs to be undertaken outdoors,despite the open air natural conditions posing many problems to the visual recognition.For example,the working environment is complicated;natural light changes greatly with time,and there are certain differences in the intensity and color of the light;the field images obtained by the plants in different growth periods have different illumination conditions,and color varies greatly;the occlusion overlap of the targets in the field of machine view produces various forms of shadows.The uncertainty and ambiguity of natural scenes in the agricultural field make image understanding a very challenging task.Therefore,the impact of illumination problems must be considered when performing identification tasks to improve efficiency,robustness and accuracy of machine vision recognition in agriculture.This paper aimed at achieving the recognition and counting tasks involved in the agricultural production process.For the first time,the group pixel algorithm has been regarded as the basic unit of image processing,combining the superpixel algorithm and the edge detection algorithm,and introducing a‘customized’ affinity matrix and ultra-metric contour map.The novel method was used in shadow detection research to construct a recognition classification model based on machine learning,which provided technical support for promoting the wide application of automatic recognition and counting based on machine learning in agriculture.The main research contents and conclusions included:(1)The FSLIC superpixel algorithm was developed in order to improve the efficiency of the algorithm and meet the real-time requirements in the recognition task of the picking robot vision system.Firstly,the image was transformed into CIELAB colour space.Norm-transformation was then applied to the distance measurement method of the original SLIC algorithm to obtain a unified mode of related parameters.In the process of allocating all pixels to the cluster center and updating the cluster center(in the iteration step),the Cauchy-Schwarz inequality and pixel norm based on weighted L2-norm theory were used,to build the condition of predict candidate cluster elimination.After these procedures,the redundant computation was reduced,and the segmentation results were optimized,and a large number of cluster searches caused by calculation of distance and comparison were eliminated.The analysis of FSLIC was based on apple images.The evaluation index included under-segmentation error,boundary recall rate and algorithm efficiency.The feasibility of image segmentation based on superpixel method in the natural illumination scene was verified.The superiority of FSLIC algorithm was also illustrated in the experiments,i.e.,the FSLIC algorithm kept the good image boundary recall,and made the superpixel segmentation speed increased by nearly 1.2 times than the original SLIC algorithm,which provided a good solution for the current problem of time consumption in the field of image processing.(2)In order to improve the overall segmentation performance and overall recognition accuracy of the visual system,FDS supeipixel algorithm and FDSE superpixel segmentation algorithms were proposed,i.e.,combined FSLIC with SEEDS algorithms to form FDS supeipixel algorithm;combined FDS with HNE edge to form FDSE supeipixel algorithm.Firstly,the HNE algorithm was used to mark the edges(denoted as edge probability map E)in the entire image to initialize the seed point;then the FSLIC algorithm and the SEEDS algorithm were applied to assign labels for each pixel(denoted as EF and Es);EF and Es were forced weighting to generate "strong boundarys",which was coincident with E,F and S "soft boundarys",which were not coincident with E;re-divide the F soft boundary based on the energy function to generate the second "strong boundary",which was coincident with the F soft boundary,otherwise labled as the weak edge.The same procedure with S soft boundary,re-divide the S soft boundary based on colour distance to generate the second "strong boundary",which was coincident with the S soft boundary,otherwise labeled as the weak edge.Finally,the weak boundary was deleted and then generated FDSE superpixel.Without the HNE-based edge marking step,the FDS superpixel was generated.Tests were carried out using small size sugarcane aphids(maximum body length of sugarcane aphid was less than 2 mm).The comparison experiments were conducted considering the quality of segmentation,the efficiency of operation,the influence of natural illumination(e.g.,high light condition,diffusive light condition,weak light condition and direct sunlight condition)and the influence of target arrangement.The results showed the feasibility of the FDS algorithm and FDSE algorithm,FDSE algorithm was thus the proper method concerning the variable light conditions in the field,providing a noval method to improve the robustness of recognition.(3)In order to reduce the impact of the shadow problem in the recognition algorithm,a new shadow detection algorithm based on UCM was proposed.The duality between the closure and the area hierarchy were defined;the hierarchical segmentation theory was applied to extract the key contours of the image;the multi-level key contour map was converted into the over-metric contour map UCM by combining the direction watershed algorithm.The parameters of each UCM feature information were optimised,and the affinity feature matrix was used to weight the UCM contour feature information.After creating the UCM map,an SVM classifier was trained to find the shadow edges from the UCM map.The detailed process is as following,using the known shadow masks marked manually,positive(negative)training examples(i.e.,edges on the UCM contour)were obtained.The positive examples were the shadow edges inside the same object,the negative examples were the others.The input features of the classifier were the affinity matrix and the corresponding superpixels’distance to the UCM contour.The output labels were the shadow edges and non-shadow edges.Thereby an image marked with a shaded area and a non-shaded area was obtained.The apple images were treated as the targets to take the comparison experiments with three classical algorithms using the receiver operating characteristic curve(ROC curve),balance accuracy(BA),improved segmentation accuracy(MSA)and relative segmentation region error value(RSAE)as evaluation criterion.Results showed that under the natural-light conditions,the proposed shadow detection algorithm based on UCM had better performance in accuracy,the lower RSAE(10.3%)in shadow recognition task,and higher segmentation accuracy(the MSA was improved in 10.1%).(4)A novel identification and counting system for natural light conditions was developed.Firstly,SVM classifier was trained:the SUN saliency algorithm was used to find the target region of interest,and then the image was divided into several sub-images for parallel processing;the original image was segmented by applying the FDSE super-pixel algorithm.The superpixel block generated after segmentation was divided into a positive sample and a negative sample,the positive sample was the target superpixel block,the negative sample was the remaining superpixel block,and the uniform LBP feature(uniform local binary pattern feature)and color feature were extracted respectively for the positive and negative samples.The dimensions of the features were reduced based on principal component analysis theory,and the features were normalized simultaneously;the model was obtained after the training step,wherein the kernel function type was set to RBF kernel,the gamma value of the kernel function was 0.03125,and the penalty function C was 32;the image regarding the superpixel block as the basic unit was tested,wherein the input features were texture features and color moment features of each superpixel,and the output tags were target superpixel blocks and else,other superpixel blocks.Sugarcane aphids were selected for recognition targets.The performance of the recognition and counting system was based on five aspects,including F-score,correct identification accuracy(IRC)and counting accuracy performance(CRC),and sugarcane locust identification error(IE)and counting error(CE).The results suggested that F-score of group-pixel algorithm(0.82-0.97)was higher than per-pixel algorithm(0.6-0.76),the recognition error was less than 0.24,and the counting error was higher 0.84 in all studied density levels.The proposed identification and counting system for complex natural lighting conditions can provide important reference and technical support for efficient pest management. |
