Abstract: This work presents a vision-based approach for detecting plant stems by separating them from surrounding elements such as leaves, soil, and background noise. An expanded dataset of early-stage corn plant images is utilized to train the model. The system detects the vertical plant structure and generates corresponding stem bounding boxes. It also evaluates the suitability of microprocessor-based platforms for efficient and portable execution of signal processing and machine learning techniques, using a Raspberry Pi 5 for deployment. The trained model achieved a mean Average Precision (mAP) of 0.777 at an IoU threshold of 0.50 and 0.346 over the IoU range of 0.50–0.95. To assess real-time applicability in agricultural settings, inference latency and CPU utilization were evaluated using field-recorded video sequences across multiple model formats. Quantized TensorFlow Lite (TFLite) implementations, including full integer and standard integer variants, achieved an average inference time of approximately 43 ms, corresponding to ~18 FPS including processing overhead. Other TFLite configurations (float16, float32, and int8) showed an average CPU usage of ~35%, representing a 42% reduction compared to the baseline PyTorch model. Additionally, inference performance and CPU utilization were analyzed for TorchScript, ONNX, OpenVINO, and NCNN frameworks.