Fruit Object Detection Research In Natural Environment Based On Deep Learning

With the rapid advancement of agricultural modernization,fruits can be picked automatically through the detection of fruit by computer vision technology.However,the traditional object detection algorithm is in low detection accuracy because it requires manual design features,and the features designed by hand are not universal.In recent years,with the growing maturity of deep learning technology,the detection accuracy has been greatly improved and deep learning-based object detection algorithm has gradually become a hot research topic However,fruits are different in shape and size in the natural environment,and are often shaded by leaves or overlapped with each other,thus affecting the detection accuracy to some extent.Therefore,this paper is aimed at improving the accuracy of different sizes of fruits detection and occlusion fruits detection through the object detection algorithm Faster R-CNN based on candidate region.The main research work of this paper is as follows:Firstly,aiming at the problem that fruit detection image training needs a large number of real samples in the natural environment based on deep learning,affine transformation is carried out on the basis of the acquired data set,so as to enhance the data and improve the robustness of the training model.Secondly,aiming at the small object easy to be missed in the fruit detection process,this paper proposes an improved multi-scale fruit detection algorithm.by redesigning the ratio and scale of candidate box,fusing the feature graphs of different levels and using multi-scale feature graphs for feature extraction.The experiments show that the average accuracy of the improved multi-scale fruit detection algorithm reaches 83.65%,which improves the overall detection accuracy and reduces the detection of small object.Finally,though the detection effect of Faster R-CNN algorithm is well improved,the detection effect of mutual occlusion of fruits in natural environment or occlusion by leaves is still not ideal,therefore,this paper proposes a fruit detection algorithm based on occlusion influence,the local context information is introduced in the feature extraction of the candidate region and improve the limitation of Non-Maximum Suppression algorithm.The experiments show that the average accuracy of the fruit detection algorithm reaches86.7%,which enhances the ability to deal with object occlusion and reduces the rate of fruit miss detection.