A novel self-attention guided deep neural network for bruise segmentation using infrared imaging

Fruit quality inspection stands as a critical pillar in the overarching framework of food safety and security. Its role in protecting public health, building consumer trust, and ensuring regulatory compliance cannot be overstated. As the industry grapples with challenges posed by globalization, technological advancements, and emerging threats, continuous innovation and collaboration between stakeholders become imperative. This introduction serves as a foundation for a deeper exploration into the dynamic world of fruit quality inspection, a field that constantly evolves to meet the demands of a changing global landscape. One of the most important steps in determining the quality of fruits is to identify and localize the bruises on apples. In contrast to the digital imaging for bruise detection, Infrared imaging technology has progressed the research for quality inspection of fruits. Despite the fact that convolutional neural networks are leading the way in image segmentation, standard models still have certain shortcomings for segmenting the bruise regions in infrared images due to the absence of any sharp boundaries and low contrast of the infrared images. Despite this fact, the paper proposes a novel multi-scale encoder-decoder architecture for bruise segmentation from infrared images thereby abstracting the contextual local features of the bruise regions through the use of guided self-attention mechanisms. Due to the non-availability of any publicly available fruit bruise dataset using infrared imaging modality, we have experimented on “TU-IR Apple Image Dataset” comprising infrared images of apple bruises. Experiments reveal that our proposed segmentation framework achieved significant performance for localizing bruises in apples with respect to the state-of-the-art competent methods with an average dice similarity coefficient of 0.8990.