Optimized Apple Recognition By Intelligence Computing Application In Harvesting Robot

Apple harvesting is a kind of high strength intensive labor and China has the largest apple planting area and yield in the world. With the change of agricultural labor structure and the development of artificial intelligence, the harvesting efficiency needs to be improved to ensure apple harvesting timely. Therefore, how to realize the automatic harvesting becomes a significant problem. The most of the existing apple harvesting robot is still applied to the laboratory research due to the low efficiency. And the objective to improve the harvesting efficiency can be achieved from two aspects:on one hand, improve the performance of the harvesting robot; on the other hand, extend the operation time, and realize the round-the-clock work. The reliability and instantaneity of the harvesting robot are directly restricted by the accurate recognition efficiency of target fruit which has become the bottleneck of visual technology.The research is carried out around the accurate recognition of target fruit to improve the efficiency of apple harvesting robot, including the night image acquisition of the apple, the analysis and de-noising of the image; the segmentation and feature extraction of the image; the establishment of recognition model, and so on. Several different algorithms for the de-noising of night image, the recognition target fruit by genetic neural network, special sample modeling are discussed in this study. The main research contents include the following aspects.Firstly, the fixed mark is adopted to capture the image of natural light and night with artificial light source. Based on the analysis of the image color, it can be concluded that contrast between RGB color components of night image is much obvious than the natural image. And the night image with incandescent lamp is more close to the natural light image. The night image is obscure and has the salt and pepper noise from the intuitive visual observation. While by the analysis of the image subtraction, it is indicated that the noise type of night image is mixed which is based on the Gauss noise and with some salt and pepper noise.Secondly, three intelligent optimization image de-noising algorithms have been established aiming at the noise problem of night vision image. The wavelet de-noising algorithm based on fuzzy threshold (F-WT) which is used to optimize the potential risks of wavelet threshold.and the relative peak signal-to-noise ratio (RPSNR) of the got low noise image will be increased by 19.69%; The combination de-noising algorithm based on WT and independent component analysis (WT-ICA) will reach a 29.94% growth of the RPSNR of the optimized low noise image; The ICA de-noising algorithm is based on particle swarm optimization (PSO-ICA) which is used to solve the separation matrix aiming at the operation efficiency of ICA, and the RPSNR of low noise image increased by 21.28% after optimization. Through comparison, the WT-ICA algorithm has the strongest ability of noise reduction but with low efficiency. While PSO-ICA algorithm has an obvious increase of efficiency with the similar level of noise reduction. Both the original images and low noise image have the highest RPSNR under incandescent light at night, so the incandescent lamp is selected as auxiliary light source during night work.Thirdly, in the process of image segmentation, the operating efficiency of the PCNN is slightly higher than the K-means clustering with both good segmentation results under Lab color space. According to the characteristics of the target fruit,the 6 color features, R, G, B, H, S, I and 11 shape features, circle variance,elliptical variance, density, circumference-to-area ratio, Hu invariant moments are extracted which can sufficient represent the target’s characteristics.Fourth, the classifier is designed according to the extracted feature vectors. RBF and Elman neural network (NN) are introduced into target fruit recognition by using genetic algorithm to overcome their inherent defects. The connection weights and network structure evolution are adopted simultaneously in this optimization to establish GA-RBF-LMS and GA-Elman optimized neural networks respectively. On the basis of the results of UCI data set simulation and apple recognition experiment, the operation efficiency and recognition precision of the two recognition algorithms is greatly improved and also the generalization ability. The overall recognition rate reaches 95% of target fruit. With comparison of the two recognition models, GA-RBF-LMS has the advantage of high operation efficiency, but the recognition precision is a little low, however the GA-Elman is opposite.Fifth, in view of the large sample problem in the process of modeling, the partial least squares (PLS) dimension reduction algorithm is adopted to release the high dimension of the sample to obtain more meaningful low dimensional data. Hierarchical cluster analysis (HCA) is used to realize the horizontal dimension reduction. The similar samples are classified as a sub-class, then each sub-class is modeled by neural network to build PLS-HCA-NN large sample classification algorithm. According to the UCI test data, it is concluded that the operation efficiency, recognition precision, and generalization ability of the two classification algorithms are greatly improved.The results of this study lay a powerful theoretical foundation for the improvement of vision system and harvesting efficiency of apple harvesting robot. Furthermore, the established intelligent optimization algorithm is also worth further application in other areas.