Deep-CofeNet: Deep Learning Based Dual Attention MobileNet for Classification of Coffee Bean Defects

Abstract Coffee bean disease must be detected earlier before roasting for good production quality. Manual screening and sorting are currently utilized in the field coffee bean industry to detect defects. In this paper, a novel Deep-CofeNet is introduced for various coffee bean defect identification to increase the production quality. Initially, the coffee bean images are gathered from Specialty Coffee Association of America (SCAA) standard lists level for classifying the coffee bean defects. Circular Bounded Inpainting Generative Adversarial Network (CBI-GAN) is used to generate the triple views (left, right, back) of a coffee bean based on its front view. Triple-side sampling (TSS) block combines the characteristics collected by the left, right and front views of the coffee bean to generate the back view for improving the quantity of the training data samples. Afterwards, the front and generated triple view of test images are pre-processed with Contrast stretching adaptive bilateral (CSAB) filtering for eliminating the noise artifacts. Finally, Dual Attention MobileNet (DA-MobileNet) is used to classify the defects into good, sour, immature, insect bite and withered from the multiple views of the coffee bean. The efficiency of the proposed Deep-CofeNet was evaluated with accuracy, specificity, precision, F1 score, recall, and circular consistency loss. The proposed Deep-CofeNet attains the highest accuracy of 98.96% using the publicly available dataset. From the analysis, the proposed Deep-CofeNet increases the average accuracy by 0.87%, 1.93%, 12.05% and 5.68% for Multiscale defect-detection deep-learning model, Machine learning model, Transfer learning networks and Convolutional neural network respectively.