Harvesting damage detection in maize ear harvesters based on the Fused Inverted Bottleneck Attention U-Net

Abstract

Improper operating parameters in maize ear harvesters can cause physical damage to the ears, resulting in grain loss. To address the lack of real-time decision-making support for correcting improper operating parameters in maize ear harvesters, we designed a Fused Inverted Bottleneck and Attention mechanism combined U-Net model (FIBA U-Net) for real-time detection of maize ear damage. The encoder of FIBA U-Net replaces the first eight blocks of EfficientNet-B3 with Fused-MBConv, which enhances the model's feature extraction capability and significantly improves detection accuracy. The CBAM module is applied to the first two skip connections in the decoder of FIBA U-Net, effectively enhancing features in maize ear damage regions while suppressing irrelevant information such as lighting variations and background noise. We compare the performance of FIBA U-Net with Eff U-Net, ResNeSt50, ConvNeXt, and ViT in segmenting the complex boundary regions of ear stacking within the maize ear collection chamber. FIBA U-Net achieves the best performance across all evaluation metrics. The method achieved an overall accuracy of up to 98.61% in field tests. The damage detection rates for the bottom of the grain collection box differed from the actual damage rates by only 0.31% under partial coverage and 0.42% under full coverage, respectively. Experimental results demonstrate that the model significantly improves accuracy, making it suitable for deployment in the complex operating environment of maize ear harvesters. It provides valuable support for parameter optimisation of harvesting machinery and maize ear grading, and demonstrates strong potential for practical application and large-scale deployment. • Novel high-precision approach for maize ear harvest damage detection. • Real-time ear damage detection via TensorRT and INT8 Optimisation. • Fused-MBConv combined with CBAM enhances complex area detection. • The method sustains high performance even under varying light intensity.

Abstract from the paper’s OpenAlex record; open copy licensed cc-by-nc-nd. Read at the source ↗

The text the model read, with the source record’s inline markup (italics, sub- and superscripts) rendered and its character entities decoded; quote offsets refer to the raw text.