Abstract: To address the challenges posed by traditional convolutional neural network (CNN) algorithms in peanut pod appearance recognition, such as high memory consumption, computational complexity, and difficulty of deployment on resource-constrained edge devices, this study proposed an efficient, lightweight, and real-time identification model named PPINET (peanut pod identification network). Designed specifically for embedded systems, PPINET achieved high accuracy and low latency while significantly reducing computational overhead, making it highly suitable for intelligent agricultural applications. The architecture of PPINET integrated depthwise separable convolutions and inverted residual blocks, which effectively reduced both the number of parameters and floating-point operations (FLOPs), without compromising the model’s ability to extract discriminative features from peanut pod images. This lightweight backbone enabled the model to run efficiently on low-power edge devices. To further enhance feature extraction, the model incorporated a lightweight attention module, the multi-query attention (MQA) mechanism. Specifically optimized for embedded deployment, MQA strengthened the network’s focus on key features, thereby improving classification accuracy and robustness under variable conditions. To adaptively optimize the model structure for deployment environments with strict resource constraints, this study adopted TuNAS, a scalable and easy-to-tune framework for efficient neural architecture search (NAS) with weight sharing. Based on TuNAS, the study designed a convolutional unit called Tun, which integrated multiple configurable inverted residual blocks and allowed dynamic architectural adjustments according to application needs. This design ensured that PPINET remained adaptable and efficient across various hardware platforms. Several training strategies were employed to further enhance model performance. A refined image preprocessing pipeline was introduced, which included contour extraction and precise cropping to mitigate distortions caused by traditional image scaling. This process significantly improved the quality of training samples and enhanced the model’s generalization ability across different peanut pod appearances. In addition, a cosine annealing learning rate schedule was applied to dynamically adjust the learning rate, accelerating convergence while avoiding suboptimal local minima. For deployment, the model was quantized into the RKNN format, enabling hardware-level acceleration on a Rockchip RK3588 platform equipped with a neural processing unit (NPU). The quantization process significantly reduced memory usage and inference latency while maintaining classification performance. The final model size was only 1.85 MB, with 0.49 million parameters and a computational complexity of just 0.30 GFLOPs, making it highly suitable for embedded environments. Experimental results demonstrated that PPINET achieved an outstanding classification accuracy of 98.65% in peanut pod recognition tasks. Further comparative evaluations showed that the inclusion of the MQA module improved recognition accuracy by up to 2.34% over conventional attention mechanisms such as SE (squeeze-and-excitation) and CBAM (convolutional block attention module), highlighting its superior performance-to-efficiency trade-off. In deployment tests, PPINET achieved an inference speed of 321 frames per second (fps) on the RK3588 development board, significantly outperforming popular embedded systems like the Raspberry Pi 4B and Jetson Nano. This real-time processing capability enabled PPINET to be effectively used in automated peanut pod sorting systems, where rapid and accurate classification was essential for improving agricultural productivity. The model’s compact architecture and hardware-accelerated design made it particularly well-suited for smart farming applications requiring low-power, real-time AI solutions. In conclusion, PPINET successfully addressed the technical bottlenecks of deploying CNN-based recognition models on embedded devices by combining a lightweight and efficient network design, a scalable NAS framework, optimized attention mechanisms, and hardware-friendly quantization techniques. This work provided a practical and reliable solution for real-time agricultural product identification and paved the way for future innovations in edge AI applications for precision agriculture.