TOPIC: Impact of FP16 Quantization on MobileNetV3Large Performance for Mung Bean Defect Classification.

In resource-constrained environments, deploying deep learning models for real-time image classification poses significant challenges due to limited computational power and memory. Existing solutions often rely on full-precision (FP32) models, which are computationally expensive and impractical for embedded systems. This study addresses the problem of efficient deployment of deep learning models by evaluating the impact of 16-bit Floating Point (FP16) quantization on the performance of MobileNetV3Large for mung bean seed defect classification. The proposed solution targets the limitations of current approaches, which offer high accuracy but at the cost of large model size and slow inference speeds. A dataset comprising 6,598 high-resolution images was constructed, with samples classified into five defect categories: broken, immature, infected, normal, and rotten. The baseline FP32 MobileNetV3Large model achieved a test accuracy of 94.85% with a model size of 16.2MB and an inference speed of 3.5 frames per second (FPS). After applying FP16 quantization, the model size was reduced to 8.27MB and inference speed increased to 8 FPS. This demonstrates a significant improvement in memory and speed efficiency. Although there was a minor accuracy drop to 93.86% (a reduction of 0.9%), the trade-off is acceptable for real-time applications on embedded platforms. These findings highlight the practical advantages of FP16 quantization for deploying lightweight yet accurate deep learning models in resource-constrained environments. The results support its viability for real-time agricultural applications such as automated seed sorting.