FruCapsNet: A shuffled attention based capsule network for multi-fruit quality assessment
摘要
With the increasing global consciousness in the nutritional impact of food consumption, there is a growing need to ascertain the quality standards of fruits or vegetables. Precise and timely quality assessment of fruits or vegetables offers substantial economic advantages to agronomists. It also provides agricultural support by supplying farmers with crucial assessments that enable fair compensation for harvests and higher-quality goods. A significant technological gap in monitoring the quality evolution of fruits within the food supply chain is leading to substantial fruit wastage. Several cutting-edge technologies, such as computer vision, machine learning, and deep learning, have been extensively applied in various agricultural domains, particularly in quality estimation, to address the food security issue. Recently, deep belief classifiers have facilitated promising outcomes, showing higher accuracy than conventional methods. However, the existing deep models for assessing the quality of fruits and vegetables are translation invariant and sensitive to the captured image orientation. Moreover, the existing methods for the fruit quality assessment are fruit-specific and fail to perform on multiple fruits and vegetables. To address this concern, this paper presents a novel method, FruCapsNet, which leverages the strength of capsule network with a shuffle attention mechanism to identify multiple fruit quality effectively. The proposed method utilizes the convolution layers for low-level feature extraction, a shuffled attention block for selective feature maps, and a capsule module for capturing hierarchical feature map relationships. The performance of the proposed method has been validated using the publicly available VegNet dataset, which contains four different fruits with multiple qualities. The experimental result showed that FruCapsNet overshadows the existing state-of-the-art deep network models in terms of accuracy, F1-Score, precision, recall, MSE, and specificity. The proposed method has achieved the highest testing accuracy of