Segmentation and Classification for Bruise Severity Detection Using Infrared Imaging

Identifying bruises in apples is one crucial step in defining the quality of fruits. The quality of apples available for sale might be significantly improved by having a detection system that can identify and remove damaged apples, which could then benefit the entire fruit economy. In this paper, we proposed a framework for classification of bruises in infrared images of apples in three major categories (i.e., slight, moderate, and severe bruises) depending on the analysis of the segmented bruised regions. For segmentation of the bruise regions, an encoder-decoder network-based segmentation method is incorporated in our proposed framework. Also, the influence of various backbone convolutional neural networks (CNNs) on encoder-decoder framework for precise bruise regions segmentation from the holistic infrared images is also investigated. After that, morphological shape features are extracted from the bruise regions for classification of bruises in apples using the well-known machine learning classifiers. The entire experiment is conducted in our own created dataset comprising of infrared images of fresh and bruised apples. Experiment results reveals that precise segmentation of the bruise regions and extraction of morphological shape features from the segmented regions can well predict the severity of bruises in apples (i.e., in three major categories including slight, moderate, and severe bruise) with an average accuracy, specificity, and sensitivity of 90.96%, sensitivity 91.37%, and 90.71% respectively.