Co-occurrence patterns based fruit quality detection for hierarchical fruit image annotation

Abstract

Automatic image annotation is a method of assigning caption to images that provide some convenient way to index, retrieve and handle a large amount of data objects. It focuses on recent agricultural automation applications; it finds potential in classification along with contextual labeling of the involved objects or detailing based on its statistical properties on fruit categories. However, producing hierarchical labels provide details of a particular fruit subcategory. This paper proposes fruit annotation in a broad sense along with its hierarchical features that can be narrowed down to inherit, further achieving fruit classification into binary or multiple classes indicating subcategories of that fruit. The fruit objects within images are measured to its actual size in the required units. The classification is also used for identifying true color, texture, size, deep features and shape based on the ratio of major to minor axis helpful for fruit gradations. The co-occurrence patterns are obtained based on the visual features of the selected fruit. This is useful for finding the fruit quality categories and combined properties that are used to form the co-occurrence patterns. These patterns are further used by the classifier for fruit annotation. The evaluation of the performance is carried out using the F1 score, accuracy, precision, recall and G-measure. The results show that the co-occurrence pattern with SVM provides an overall accuracy of 97.3% and 97.2% for grape and mango fruit subcategories. The comparative results are obtained to cross-check with the subjective evaluation of gradation validated by local farmers.

Keywords

• Machine learning
• Co-occurrence pattern
• Fruit annotation
• Hierarchical labels