Compute-Aware Evaluation of Severity Conditioning and Lightweight Architectures for Real-World Image Deraining

Abstract

Image deraining for embedded perception must balance restoration quality against complete inference cost. Yet density-conditioned methods require annotations unavailable for captured rain, while efficiency studies often omit conditioning-signal cost. This study examines whether label-free severity conditioning justifies its end-to-end cost and whether complexity measures predict realized latency. Under a common L₁ protocol, we compare a PReNet baseline with scalar feature-wise linear modulation, spatial severity modulation (SSM), a matched-capacity constant-map control, and two lightweight architectures. Zero-, shuffled-, and constant-prior interventions test SSM’s dependence on prior presence, image correspondence, and spatial structure. Six models were trained for 100 epochs on 10,000 LHP-Rain patches using three seeds and evaluated on the 1,000-image test split. SSM achieved 31.775±0.109 dB, compared with 31.741±0.095 dB for Base and 31.754±0.102 dB under zero-prior replacement. A seed-level equivalence analysis based on paired image differences supported practical equivalence between SSM and Base (95% CI: ?0.026 to 0.094 dB; margin fixed before primary analysis: ±0.1 dB), consistent with the half-scale experiment. The two-stage model reduced computation from 265.0 to 44.6 GFLOPs and model latency from 20.43 to 4.50 ms, but the prior generation increased total latency to 24.74 ms. CBAM-light required 14.1 GFLOPs yet recorded 33.68 ms. The label-free statistic achieved a cross-scene AUC of 0.772 on RealRain-1k, indicating moderate agreement with density labels. Thus, severity-conditioned and lightweight deraining systems should be assessed through verified prior use and measured end-to-end latency rather than accuracy, parameters, or GFLOPs alone.

ABSTRAK: Penyahhujanan untuk sistem terbenam perlu mengimbangi kualiti pemulihan dengan kos inferens menyeluruh. Namun, kaedah yang bergantung pada kekerapan memerlukan anotasi yang tidak tersedia untuk hujan yang dirakam, manakala kajian kecekapan sering kali mengabaikan kos isyarat pengkondisian. Kajian ini meneliti sama ada penentu tahap tanpa label membenarkan kos hujung-ke-hujung dan sama ada ukuran kerumitan meramalkan kelewatan sebenar. Di bawah protokol L1, garis dasar PReNet dibandingkan dengan modulasi linear mengikut ciri skalar, modulasi tahap spatial (SSM), kawalan peta tetap berkapasiti sepadan, dan dua seni bina ringan. Intervensi sifar-, diacak-, dan malar menguji kebergantungan SSM pada kewujudan, kesepadanan imej, dan struktur spatial. Enam model dilatih untuk 100 epok menggunakan 10,000 tampalan LHP-Rain dengan tiga benih dan dinilai pada 1,000 imej ujian. SSM mencapai 31.775±0.109 dB, berbanding 31.741±0.095 dB bagi Base dan 31.754±0.102 dB di bawah penggantian sebelum sifar. Analisis kesetaraan tahap biji berdasarkan perbezaan imej berpasangan menyokong kesetaraan praktikal antara SSM dan Base (95% CI: ?0.026 hingga 0.094 dB; margin ditetap sebelum analisis utama: ±0.1 dB), ianya konsisten dengan eksperimen separuh skala. Model dua peringkat mengurangkan pengiraan daripada 265.0 kepada 44.6 GFLOP dan model pendaman daripada 20.43 kepada 4.50 ms, tetapi bagi penjanaan sebelum, meningkatkan jumlah pendaman kepada 24.74 ms. CBAM-light memerlukan 14.1 GFLOPs tetapi merekodkan 33.68 ms. Statistik tanpa label mencapai AUC rentas senarai 0.772 pada RealRain-1k, menunjukkan persetujuan sederhana pada label ketumpatan. Jadi, sistem penyahhujanan ringan dan bersyarat teruk perlu dinilai melalui penggunaan yang disahkan sebelumnya dan ukuran kelewatan hujung-ke-hujung, bukan pada ketepatan, parameter, atau GFLOPs.