| Abstract: |
Wheat entering India's central pool is graded by hand. An inspector draws a working sample, spreads it on a tray, and picks out refractions one category at a time with a magnifying glass, following an order fixed by statute: other food grain first, then damaged, discolored, insect-damaged, fragments, broken, slightly damaged, and finally shrivelled or immature. The procedure is slow, it varies between inspectors, and disputes over marginal calls at procurement centres have been documented as a source of farmer hardship. This study treats that manual protocol not as a heuristic to be replaced but as a specification to be reverse-engineered. The ordered hand-picking sequence encodes a priority-encoded, single-assignment classifier each kernel is accounted to exactly one refraction and that structure maps directly onto a hardware cascade. An eight-class taxonomy was derived from the Food Corporation of India Uniform Specification and the associated Bureau of Indian Standards methods, and a five-stage field-programmable gate array vision architecture was specified around it: line-scan acquisition, programmable-logic pre-processing and segmentation, an INT8 MobileNetV2 backbone on a deep-learning processing unit, and a priority cascade that enforces the statutory ordering in fixed logic rather than in a softmax layer. Projected performance was derived from published accelerator benchmarks across six platforms. The design sustains 340 kernels per second at 4.08 ms end-to-end latency on a Zynq UltraScale+ ZU9EG, reducing a 100 g working sample from roughly twenty-five minutes of hand-picking to about seven seconds. |
