基于多尺度超分辨率生成对抗网络的水稻病害识别方法

Rice Disease Recognition Method Based on Multiscale Super-resolution Generative Adversarial Network

摘要:

针对水稻病害识别过程中受采集条件限制导致图像分辨率较低、病斑细节缺失，从而影响识别精度的问题，本文提出了一种基于多尺度超分辨率生成对抗网络的水稻病害识别方法（ESRG_E）。该方法在超分辨率重建网络(ESRG)中，首先，设计了动态多尺度残差嵌套块（DMS-RRDB），其中动态多尺度机制增强网络对不同尺度病斑结构的表征能力，残差嵌套结构提升高频细节的重建能力;其次，设计并行双窗口自注意力模块（PDSA），通过设置窗口不同尺寸，以增强病斑区域特征的聚焦能力和空间结构一致性，从而突出关键病斑结构;此外，引入双感知损失函数（DP Loss），以强化重建结果的判别性约束，缓解传统损失函数导致的过度平滑问题。最后，将超分辨率重建网络与识别性能较优的EfficientNet分类网络相结合，以提升病害识别精度。试验结果表明，在图像超分辨率重建方面，峰值信噪比（PSNR）和结构相似性指数（SSIM）平均分别提升1.43 dB和0.035;将超分辨率重建网络与多种分类模型结合后，其准确率均高于低分辨率输入，平均提升6.3个百分点。在模型对比试验中，本文方法水稻病斑识别准确率较其他模型提升1.28个百分点，最终达到97.6%。本文方法提升了低分辨率水稻病害图像重建质量与识别性能，为复杂田间环境下水稻病害智能监测提供了技术支撑。

Abstract:

Aiming at the problem that low image resolution and missing lesion details caused by limited acquisition conditions in rice disease recognition degraded the recognition accuracy, a rice disease recognition method was proposed based on a multi-scale super-resolution generative adversarial network (ESRG_E).In the proposed method, a super-resolution reconstruction network (ESRG) was firstly constructed by designing a dynamic multi-scale residual-in-residual dense block (DMS-RRDB), in which the dynamic multi-scale mechanism enhanced the representation ability for lesion structures at different scales, while the residual-in-residual architecture improved the reconstruction of high-frequency details.Then, a parallel dual-window self-attention module (PDSA) was introduced, where different window sizes were adopted to strengthen the focus on lesion regions and maintain spatial structural consistency, thereby highlighting key lesion structures.In addition, a dual-perceptual loss function (DP Loss) was employed to enhance discriminative constraints on the reconstructed results and alleviate the over-smoothing problem caused by traditional loss functions.Finally, the super-resolution reconstruction network was combined with the EfficientNet classification network with superior recognition performance to further improve disease identification accuracy.Experimental results showed that, in terms of image super-resolution reconstruction, the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) were improved by an average of 1.43 dB and 0.035, respectively.When combined with various classification models, the reconstructed images achieved higher recognition accuracy than low-resolution inputs, with an average improvement of 6.3 percentage points.In comparative experiments, the proposed method improved rice lesion recognition accuracy by 1.28 percentage points over other models, finally reaching 97.6%.The proposed method significantly enhanced the reconstruction quality and recognition performance of low-resolution rice disease images, providing technical support for intelligent monitoring of rice diseases in complex field environments.