DEF-dhif: Dynamic High-frequency Convolution for IR Small Target Detection

ANIMA defense module implementing DHiF (Dynamic High-frequency Filter) convolution for infrared small target detection (IRSTD).

Architecture

• UNet backbone with DHiF convolution in bottleneck
• DHiF generates input-dependent dynamic operators for high-frequency enhancement
• Trained on NUAA-SIRST dataset (255 train / 85 test images)

Paper

• Title: Dynamic High-frequency Convolution for Infrared Small Target Detection
• ArXiv: 2602.02969

Exports

• model.pth — PyTorch checkpoint
• model.safetensors — Safetensors weights
• model.onnx — ONNX (opset 17, dynamic axes)
• model_fp16.plan — TensorRT FP16
• model_fp32.plan — TensorRT FP32

Usage

import torch
from safetensors.torch import load_file
from def_dhif.models.unet_dhif import UNetDHiF

model = UNetDHiF(in_channels=1, out_channels=1, base_ch=64)
state = load_file("model.safetensors")
model.load_state_dict(state)
model.eval()

# Single-channel IR image, padded to 512x512
x = torch.randn(1, 1, 512, 512)
pred = model(x)  # Binary segmentation mask

Built with ANIMA by Robot Flow Labs