Abstract: Maize ear-harvesting is the dominant maize mechanized harvesting mode in China owing to its tolerance of a wide range of kernel moisture contents. During picking, conveying, elevating, and husking, however, maize ears inevitably suffer squeezing, collision, and friction against metallic components and other ears, producing diverse mechanical damage such as gnawed ears, skin abrasions, kernel indentations, and kernel fractures. Such damage degrades storage quality and introduces hidden yield losses, yet its quantification currently relies solely on post-harvest sampling inspection, with no means of real-time in-field detection or adaptive feedback to harvester control systems. Developing accurate and timely damage recognition techniques is therefore of considerable importance for improving maize yield and safeguarding food security. Conventional approaches combining hand-crafted image features with classical classifiers such as support vector machines achieved preliminary damage recognition but lack robustness under the disordered ear stacking, variable illumination, and heavy background interference encountered inside actual grain tanks. More recent convolutional neural network methods, particularly YOLO-family detectors, have improved performance on tasks such as kernel crack identification, missing-kernel detection, and fallen-ear detection, yet they provide only bounding-box-level localization and cannot delineate pixel-level damage contours, limiting their utility for precise harvester parameter adjustment. To address these limitations, this study proposes FIAC U-net, a pixel-level semantic segmentation model built on the U-net encoder–decoder framework with two targeted improvements. First, the first six convolutional blocks of an EfficientNet-B3 encoder are replaced with Fused Mobile Inverted Bottleneck Convolution (FusedMBConv) modules, which unify the 1×1 expansion and 3×3 depthwise convolution of standard MBConv into a single 3×3 convolution. This reduces memory-access overhead and inference latency while enhancing multi-scale feature extraction. Each FusedMBConv unit further incorporates batch normalization and a Squeeze-and-Excitation channel-attention mechanism to adaptively recalibrate channel responses. Second, a Convolutional Block Attention Module (CBAM) is inserted at the skip connection of the deepest decoder layer. CBAM sequentially applies channel attention via shared multi-layer perceptrons on pooled descriptors and spatial attention via a 7×7 convolution on concatenated average- and max-pooled feature maps, enabling the model to emphasize damage-relevant features while suppressing background and illumination interference. The dataset was collected during the October 2024 field harvest of Xianyu 335 maize in Dunhua, Jilin Province. Fifty-five top-view RGB images of naturally stacked ears under varied illumination and piling conditions were annotated into four classes, namely intact maize, damaged maize, husk leaves, and background, by two independent annotators with consensus resolution. A five-fold nested cross-validation scheme was adopted, and data augmentation expanded the training sets to 7224 (inner) and 8832 (outer) images. The encoder was initialized with ImageNet-pretrained weights and fine-tuned with early stopping; weighted cross-entropy loss addressed the severe class imbalance of the damage category. Ablation results showed that segmentation performance plateaued at 9–11 encoder FusedMBConv blocks. Further combinatorial ablation showed that six encoder FusedMBConv blocks and one decoder CBAM module achieved near-optimal accuracy with substantially lower computational cost. This configuration was therefore adopted. Comparative experiments against ConvNeXt, ResNeSt50, SegFormer, and SegMAN under identical settings showed that FIAC U-net achieved the highest overall accuracy of 93.48%, macro-F1 of 86.23% (2.64 percentage points above the best baseline SegMAN), and a damaged-maize recall of 88.06% with an F1-score of 85.42%, demonstrating a superior balance between missed and false detections. Pixel-level error maps confirmed cleaner segmentation with misclassifications concentrated narrowly along object boundaries. The average per-image inference time was 0.19 seconds on a Jetson Xavier NX platform with TensorRT INT8 quantization, confirming suitability for real-time deployment. A generalization experiment using a different variety (Jinqing No.1) under varying operating conditions yielded a damage-class precision of 84.49%, verifying the model’s robustness and cross-variety adaptability. This work provides a technical foundation for integrating damage perception into intelligent maize harvesting systems and supporting future closed-loop sensing–decision–control frameworks for adaptive harvester parameter regulation.