Pollen recognition skips full heat conduction on every channel: 14.9% fewer parameters, 9.9% more throughput
Synopsis
Pollen microscopy recognition must separate subtle differences between species while absorbing large variation within a species, and the heat-conduction vision backbone vHeat runs a complete two-dimensional heat-diffusion transform on every channel. To cut that redundancy, part of the channels are routed through a lightweight local branch, a controlled frequency-retention constraint is applied to mid-to-late DCT responses, and both ratios are set per network stage. On the in-house Pollen149 set, parameters fall from 30.20M to 25.69M and FLOPs from 4.94G to 3.97G with a 0.19 percentage-point Top-1 drop, while throughput rises from 924.15 to 1015.27 images/s.
Interpretation
A pollen recognition model can reserve full heat-conduction computation for only a subset of channels, lowering both parameter count and floating-point work without changing the classification pipeline. The earlier vHeat backbone applied the complete two-dimensional heat-diffusion transform to every channel: a discrete cosine transform, frequency-domain modulation, then inverse reconstruction. On Pollen149, the stage-wise configuration cut parameters from 30.20M to 25.69M and FLOPs from 4.94G to 3.97G, with Top-1 moving from 96.20% to 96.01%, a gap of 0.19 percentage points.
Setting channel routing and frequency constraints separately at each network depth pays off compared with one ratio for all stages. Uniform channel routing reached 95.66% Top-1, and uniform routing plus frequency regularization fell further to 95.45%. The stage-wise configuration reached 96.01% Top-1 at 1015.27 images/s and 0.9850 ms latency, against 924.15 images/s and 1.0821 ms for the original vHeat.
The implementation splits the input feature along channels, sends one part through the full heat branch and the rest through a depthwise convolution with GELU, then concatenates outputs to keep the original channel count. The full heat branch previously carried global interaction for all channels, while the lightweight branch performs only local transformation at clearly lower cost. On the frequency side, a low-frequency retention mask is inserted after heat-diffusion modulation and before inverse transform, conservative in shallow stages to keep edges and textures and stronger in middle-to-late stages, with ratios chosen by ablation on the accuracy-efficiency trade-off.
Perspective
The results cover the in-house scanning-electron-microscopy dataset Pollen149 (29,018 images at 224×224, 51 families and 149 species) plus the New Zealand Pollen Dataset and CPD-1, in a setting of batch pollen recognition on microscopy images. The authors propose future work on improving execution efficiency of partial heat routing and frequency constraints across hardware platforms, and on more flexible channel routing and frequency constraints.
Measured speedup is smaller than the theoretical computation cut: FLOPs fall 19.6%, while throughput rises about 9.9% and latency drops about 9.0%, which the authors attribute to memory access, operator implementation and GPU utilization. Frequency constraints may still suppress useful high-frequency microscopic detail such as surface ornamentation, pores, grooves and local protrusions; across three seeds the final configuration reaches 95.90±0.11% against 95.93±0.31% for the original vHeat, comparable on average.