Research On Key Technologies Of Fruit And Vegetable Recognition Based On Multi-Instance Multi-Label Learning

Promoted by agricultural industrialization, supermarket operations for produce become the leading mode of retail terminal. A solution for produce image recognition based on multi-instance multi-label learning is proposed to address the problem of produce’s price confirmation which depends on human memory in the sale process of super market and realize intelligent identification of agricultural products. The main works are as follows:(1) Research on learning frameworks fitting for multi-class produce image classification. On the basis of comparisons between supervised learning and semi-supervised learning frameworks, multi-instance multi-label learning is determined to model the problem. The framework overcomes the inadequacy of supervised learning which can handle single label images only.(2) Research on bag generation method of multiple instance learning. Two bag generation methods are proposed in this paper. One generates fixed-shape regions as instances. By employing concept of goal-orientation, the method based on single blob with neighbors which proposed for scene image classification originally can handle object image classification now, and can effectively hit fruit and vegetables in images. Another bag generation method is based on image segmentation. Specifically, R, G, B value of pixels are taken as samples, and organized for clustering based on k-means method. Given that any clustering method should provide clustering number which is embarrassing in image segmentation, an evaluation index is introduced. On top of this, the index has been improved to be scalability which meets needs of self-adaptive segmentation.(3) Research on degenerating strategy of multi-instance multi-label learning. Compared with degenerating strategy with multiple instance learning as bridge and multiple label learning as bridge, degenerating to multiple instance learning is more in line with characteristics of the images in the research. Finally, a multiple instance learning algorithm, EM-DD, is adopted to train produce model.(4) A total of5189images are collected as the hybrid produce image library. Experiments are carried out on self-collecting image library and Amsterdam library. The results show that the proposed method can recognize multi-class produce images which are randomly positioned and captured with distractors under different illumination, the precision rate can achieve86.22%and84.04%under two method each when applied to multi-class produce images. Moreover, classification results is better than global method when applied to single category produce images, and the precision rate can achieve95.45%. In general, it is feasible of utilizing MIML-based image classification method to provide auxiliary decision for automatic produce sales in supermarket.