<![CDATA[Analyzing corn cobs in precision agriculture presents unique challenges due to diverse shapes, textures, and dynamic lighting, which often degrade the boundary precision of conventional algorithms. To address this, this study evaluates the semantic segmentation performance of the DeepLabV3+ architecture by comparing three residual network backbones: ResNet50, ResNet50V2, and ResNet101. The experiments utilized a public dataset comprising 1,350 corn cob images, strictly partitioned into 800 training, 450 validation, and 100 testing images with normalized pixel values to ensure rigorous testing conditions. The results reveal a distinct operational trade-off. ResNet101 achieved the highest segmentation accuracy with a peak mean Intersection over Union (mIoU) of 96.45%, making it ideal for high-fidelity offline analysis. Conversely, ResNet50V2 delivered the best computational efficiency at 9.49 frames per second, perfectly suited for resource-constrained edge devices. However, the study acknowledges a primary limitation regarding a severe class imbalance that slightly degrades the segmentation of minor objects, highlighting the continuous trade-off between absolute accuracy and computational latency. Ultimately, this research provides a vital and urgent contribution to computer vision by delivering an optimized framework that explicitly bridges the gap between complex deep learning models and practical, edge-based production pipelines for automated precision farming.]]>
