<p>As a vital raw material in agricultural processing, the quality of peanut kernels critically determines the economic value and industrial applicability of derived products. Traditional sorting methods, which rely heavily on manual visual inspection, are inefficient and error-prone, making them unsuitable for the high efficiency and standardization required in modern agriculture. To overcome these limitations, this paper presents ADDF-YOLO, a lightweight model for accurate peanut kernel quality detection. The model introduces the ADown module to replace standard convolutional downsampling, thereby enhancing the retention of semantic and spatial features. The neck network employs DySample as a learnable upsampling operator to improve small-object detection capability. Concurrently, the detection head is replaced with Detect_Efficient, reducing computational complexity while maintaining high precision. Furthermore, the Focaler-CIoU loss function is adopted to supplant CIoU, intensifying the model’s focus on hard samples during training. To evaluate the proposed model, a dedicated PQT peanut dataset was constructed. Experimental results demonstrate that, compared to the baseline YOLOv11n, the proposed ADDF-YOLO reduces the number of parameters and computational cost by 28.3% and 34.9%, respectively, while achieving improvements of 2.0% in mAP@0.5, 1.3% in mAP@0.5:0.95, 1.5% in precision, and 2.2% in recall. The optimized architecture achieves a total processing speed of 175 FPS on edge devices, demonstrating strong potential for practical deployment in industrial inspection scenarios. These results verify the algorithm’s effective balance between accuracy and efficiency, providing a viable solution for real-time peanut kernel quality detection in agricultural applications.</p>