Research Of Vegetable Image Recognition Based On Deep Learning

Intelligent management has gradually been achieved promising results in varies area of our daily life with the rapid development of science and technology recent years.While referring to the vegetable recognition area,it has not been realized fully intelligent when it comes to the transaction of vegetable products in supermarkets or vegetable farms as well as the management of raw vegetable materials in smart refrigerators.Image recognition technology is in need to solve this problem.Different from other domains,the research of vegetable image recognition seriously lacks the support of specific image datasets yet.This paper introduces novel vegetable image datasets after eliminating several ambiguous interference images from Veg Fru,a domain-specific dataset consisting of fruits and vegetables.Considering that the existing literatures generally construct mix-grained vegetable datasets for researches,we separately introduce datasets under different grained.The fine-grained datasets named cabbage5,melon12,eggplant7,vegetable15,mostly belong to the same sub-categories with complex scenes and general recognition challenges in fine-grained images while the coarse-grained one named coarse-vegetable15 contains 15 common vegetables in daily life of distinguish features between classesFirstly,this paper makes researches on fine-grained vegetable datasets.Supermarket Produce Dataset,Oxford 17 Flowers are introduced for the horizontal comparison and the traditional image recognition technology as well as conventional convolution neural networks are employed to validate the effectiveness of the proposed algorithms.The former adopts a robust one based low-level feature extraction,the BOW model with SPM method and the later adopts small-scale Caffe Net and Resnet10 networks.Then a novel network named new Net is designed in view of compressing network parameters,which introduces the small sized convolution cores,minimizes application of the pooling layers and fully connected layers.Experimental results show that compared with Caffe Net and Res Net10,new Net raises recognition performance with less parameters,and respectively reaches 9.95%,24.45%,36.99%,25.80% higher than the traditional technology;meanwhile,aiming at the distinguishing difficulty in fined-grained images,a network with layer-skipping is proposed.This network connects low-level features to another network branch composed of smaller convolution kernels and combines Center Loss to fuse the low-level features and high-level features to enables network to learn more detailed information.Experimental results prove the effectiveness of the proposed network structure with greater accuracy than new Net at 76.73%,66.28%,66.65% and 61.37%.Finally,this paper makes researches on the coarse-grained vegetable image set.Since coarse-vegetable15 is a small-scale data set,the training may be too insufficient to lead the over-fitting problems by the direct application of the convolution network model.This third part proposes a recognition technology based on transfer learning to solve the shortages of labeled vegetable images.By using the pre-trained models on the large-scale data sets to transfer partial network structures and relevant parameters with initializing two additional adaptive layers,the network achieves a highest recognition accuracy of 94.97% on Caffe Net model and 96.69% on Res Net10 model under low computational cost and training time cost.