Rice seed image classification based on HOG descriptor with missing values imputation

Abstract

Rice is a primary source of food consumed by almost half of world population. Rice quality mainly depends on the purity of the rice seed. In order to ensure the purity of rice variety, the recognition process is an essential stage. In this paper, we firstly propose to use histogram of oriented gradient (HOG) descriptor to characterize rice seed images. Since the size of image is totally random and the features extracted by HOG can not be used directly by classifier due to the different dimensions. We apply several imputation methods to fill the missing data for HOG descriptor. The experiment is applied on the VNRICE benchmark dataset to evaluate the proposed approach.

PhosopNet: An improved grain localization and classification by image augmentation

“…Many computer vision papers have been gradually shifted from traditional bag of words to CNN paradigm [8] to solve object localization and classification problems in big data. Argubly, the industrial requirements concern user experience, environmental implementation and software maintenance friendliness; it was sometimes better to be implemented by histogram of gradients (HoG) [42] with traditional machine learning as a bag of words model [20][21][22], for the open-world grain inspection [23].…”

“…To do more with less data, this paper named PhosopNet proposes the image augmentation that generates the thoundsand grain data from hundred one, instead of manually labeling those ten-thoundsand small grain images. For the expansion of previous works, the computer vision applied to rice or grain problems (both bag of words [9][10][11][12][13][14][15][16][17][18][19][20][21][22][23] and CNN [24][25][26][27][28][29]) can achieve high performance by training the less labeled grain data.…”

“…Since the previous papers on grain (or rice-grain) recognition (in both bag of words [9][10][11][12][13][14][15][16][17][18][19][20][21][22][23] and convolutional neural network [24][25][26][27][28][29]) require a large number of human-annotated labels (labeled data), the proposed PhosopNet leverages the image augmentation by feature transformation to artificially generate a variety of grain instances as a larger dataset and train them to the model. Practically, a grain object that is performed by image processing operations: rotation, brightness adjustment and horizontal flipping, in order to increase a number of data, as shown in Figure 2.…”

“…Another reason was ResNet has only skip connection, while DenseNet [45] had with dense block between the layers that could easily use the feature maps from the low-level layers. According to economic condition, the grain inspection based on bag-of-words model (traditional machine learning with feature extraction) is still required [20][21][22] in the open-world industry such as iRSVPred [23]. The previous papers [29] has already showed the high correctness (higher than 0.7) in the problems: purity between glutinous and paddy grain (Phosop-1); and paddy grain grading (Phosop-2).…”

“…From the literature, all papers could be categorized by methods [8] into 2 groups: bag of words [9][10][11][12][13][14][15][16][17][18][19][20][21][22][23] and convolutional neural network (CNN) [24][25][26][27][28][29]. The former was used by early researches [30] (which require less number of labeled data [31]) that were still useful in some open-world industry [20][21][22] such as iRSVPred [23]. The latter was exponentially increased by current researches (which required large volume of labeled data [32,33]) that had already been proven to be higher performance than bag-of-words methods [34].…”