Post-harvest sorting of fresh corn based on semantic segmentation

To ensure the quality of fresh corn and improve production line efficiency, automating the identification and detection of deficiencies and diseases in corn grains is essential. This study addresses the limitations and accuracy issues among contemporary techniques in real-world production settings by developing a corn detection model known as Corn-Net, which utilizes semantic segmentation. By imitating the UNet architecture, Corn-Net incorporates a multi-scale strip convolutional attention block (MSCAB) into its encoder–decoder structure, significantly enhancing feature extraction for various sizes of corn kernels, deficiencies, and diseased areas. To mitigate the potential computational costs and efficiency reductions associated with large-scale convolution kernels, MSCAB adopts a strip convolution kernel design of 1 + K, effectively achieving a receptive field effect similar to K. Experimental results demonstrate that Corn-Net, utilizing the 1 + K design, achieves an average segmentation rate of 83.03%, an average Dice coefficient of 80.22%, and an average accuracy of 90.82%, surpassing vanilla K designs. Furthermore, this approach reduces the model latency by 41.09% and increases the frame rate, thus improving the operational efficiency compared to K designs. Corn-Net exhibits superior segmentation performance on a self-compiled fresh corn dataset compared to existing models. Incorporating Corn-Net into a unified hardware and software system offers a practical automated solution to sorting corn grades, marking a significant advance in automating and enhancing agricultural production processes.