The global issue of “garbage encircling cities” is intensifying, making intelligent waste sorting a research hotspot for tackling this challenge. However, embedded platforms commonly face the trade-off dilemma of “limited computing power - high real-time requirements - optimal recognition accuracy.” Traditional approaches struggle to meet practical demands: cloud-based architectures suffer from high latency due to data transmission, pure embedded architectures lack sufficient computing power, and cloud-edge collaborative architectures still exhibit interaction delays. This paper proposes a heterogeneous collaborative computing architecture based on FPGA-STM32. The FPGA handles image preprocessing and parallel convolution computations, while the STM32 manages fully connected layer operations and classification decisions. Concurrently, a lightweight convolutional neural network is optimized through pruning into a “single convolution layer + three fully connected layers” structure, incorporating INT16 quantization and clipping mechanisms to balance accuracy and hardware adaptability. Experiments demonstrate that the system achieves an 83.33% accuracy rate in identifying ten categories of household waste. Compared to the MATLAB platform, it accelerates inference by 15.675 times with a processing latency of only 40.004ms. The low FPGA core resource utilization enables efficient deployment in embedded waste sorting scenarios such as communities and households.