A Method Of Green Citrus Detection Based On Deep Bounding Box Regression Forest

Owing to the long-term dependence on manual production of agricultural industry in China,the problem of large labor intensity but low efficiency has directly increased the cost of agricultural production.Using machines and automatic control techniques to replace labor is one of the key ways to increase the efficiency of the agricultural industry and reduce costs.This paper mainly studies the visual detection methods of green fruits in the natural environment.The green citrus was taking as the research object in this paper.The detection of green fruits can provide important technical support for fruit yield estimation and automatic picking,which has important economic benefits and research significance.However,in the natural environment,it is difficult to distinguish the green fruits and background due to the similarity color,which has always been a difficult problem to deal with in the agricultural robots.Therefore,this paper proposed a fruit detection method based on deep bounding box regression forest.This method mainly includes feature extraction and classifier design.Based on the analysis of green citrus,this paper used three features including texture,shape and color to describe the image.In the part of texture feature,this paper put forward a boosting discriminate descriptor,which was an improve method of the CS-LBP operator.This method introduced a supervised learning method in CS-LBP binary encoding by using linear discriminant analysis and adaptive boosting.It makes the binary encoding easier to distinguish the target and background in the image.In terms of shape features,this paper compared the hough detection method with the HOG descriptor and adopt the HOG descriptor in this paper.In terms of color features,three features of better performance in various colors were chosen to describe the images through subset search method.In addition,a simple method was adopted to extract multi-scale feature of images without significantly increasing the number of image features.The method can not only extract the features of different granularity,but also reduce the correlation between the features extracted from different descriptors.The design of classifier includes model training and model inference.In the part of model training,this paper designed a kind of structured label for image patch firstly.The structured label contains the target's category information and location information.In the part of category information,a probability category label was designed to blur the categories of objects.And in the part of position information,the method of circle fitting was used to estimate the bounding box of the object.Then the Single layer regression forest was trained by residual sum of squares function to minimize class uncertainty and position uncertainty alternately.Furthermore,the output of the model was used as a new feature to series with the input features to train multiply layer regression forest and reduce the mean square error of the model output.In the prediction stage,the location of the target was voted by the system clustering method,which realized that different image patchs located different citrus.Finally,the hyper parameters were optimized by experiment,and then adopted to the model for citrus detection.The mAP of the proposed method in this paper was 87.6% and the average run time is 0.759 s per image on 800 test citrus images.