Enhancing green guava segmentation with texture consistency loss and reverse attention mechanism under complex background

Abstract

Green crops like guava, unlike the majority that can be distinctly separated from the background by color contrast, often share color characteristics with the surrounding leaves, making the accuracy of crop detection and further pixel-level prediction decrease in complex natural environment. Therefore, this work took the texture and boundaries of crops as the focus, proposed a Gabor based texture consistency loss and a reverse attention module (RAM). Meanwhile, both a receptive field module (RFM) and a mutual fusion decoder (MFD) were proposed to enhance the utilization of semantic information. Finally, a stepwise prediction refinement method with the deep prediction map as prior information was designed in the model framework, realizing a further enhancement of the inference ability. In the ablation experiments, this work verified the effectiveness of the proposed improvements step by step using Classification Evaluation Metrics and provided the visualization of the reverse attention. In the comparative experiments, this model demonstrated its advantages in contrast to state-of-the-arts such as U-Net, SETR, and SegFormer. The Acc and IoU reached 0.9954 and 0.9420, exceeding those of SegFormer by 0.0087 and 0.0119 respectively, demonstrating its application potential for agricultural robot visual systems. Moreover, to further demonstrate the inference capability of the proposed model, we conducted validation on two open-source building extraction datasets, WHU and MBD, which have similar task difficulties, and achieved significant results.

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.