Weed image classification with generative AI using latent denoising diffusion probabilistic model and Wiener filtering approach

• A novel framework combining Latent Denoising Diffusion Models (LDM) with Wiener filtering for robust weed image classification under real-field conditions. • Integration of neuromorphic spiking neural networks (SNNs) for content-based image retrieval, enhancing structural feature recognition. • Theoretical justification for Wiener filtering in the latent domain to improve texture preservation and noise robustness compared to conventional denoising methods. • Extensive experimental validation including cross-dataset generalization, t-SNE visualizations, and denoising performance comparisons with multiple baselines. • Significant improvements in classification accuracy and structural similarity metrics over traditional GAN-based and CNN-based methods. A generative framework for weed image classification has been proposed in this study based on a Latent Denoising Diffusion Probabilistic Model (DDPM) integrated with a convolutional neural network (CNN)-based U-Net architecture. The method addresses the challenge of limited labeled training data in agricultural datasets by synthesizing high-fidelity weed images that accurately reflect the visual and structural complexity of real-world field conditions. Within this framework, synthetic images are generated through a forward process that incrementally adds Gaussian noise to real images, followed by a learned reverse denoising process modeled via a U-Net-based neural network. Wiener filtering is applied in the latent domain to enhance frequency-domain consistency and suppress background noise. The synthetic images are subsequently employed for both dataset augmentation and supervised weed classification. The proposed framework was evaluated on four public datasets: DeepWeeds, soybean weeds, corn weeds, and cotton & tomato weeds. The highest performance was achieved on the DeepWeeds dataset, with an MSSIM of 0.61 (original dataset score: 0.63) and classification accuracy of 98.52% using the LDM-augmented data. The corn weed dataset yielded 0.67 MSSIM (original: 0.72) and 94.28% accuracy, followed by soybean weeds (MSSIM: 0.65, accuracy: 67.92%) and cotton & tomato weeds (MSSIM: 0.42, accuracy: 33%). The results demonstrate that latent diffusion-based image synthesis combined with frequency-aware denoising significantly improves the robustness and generalization of weed classification models under real-field variability.