Banca de QUALIFICAÇÃO: DAVI SAMUEL DE FRANÇA BRAGA

Uma banca de QUALIFICAÇÃO de MESTRADO foi cadastrada pelo programa.
STUDENT : DAVI SAMUEL DE FRANÇA BRAGA
DATE: 25/09/2026
TIME: 10:00
LOCAL: Remoto
TITLE:

Texture-Aware Pruning for Fine-Grained Visual Categorization

KEY WORDS:

Fine-Grained Visual Categorization, Neural Network Pruning, Edge Artificial
Intelligence, Additive Fusion, Computer Vision.

PAGES: 60
BIG AREA: Ciências Exatas e da Terra
AREA: Ciência da Computação
SUBÁREA: Metodologia e Técnicas da Computação
SPECIALTY: Sistemas de Informação
SUMMARY:

Fine-Grained Visual Categorization (FGVC) is a subfield of Computer Vision focused
on distinguishing subordinate classes with high morphological similarity, such as specific
species of animals, plants, or vehicle models. Unlike generic image classification, success in
this domain requires the meticulous extraction of local micro-details and subtle textures
to overcome the low inter-class variance. Although deep Convolutional Neural Networks
exhibit high performance in this task, their massive computational cost and parametric
redundancy preclude direct deployment on resource-constrained devices. In the context
of model compression via Network Pruning, the application of traditional criteria based
purely on structural magnitude, such as the L1 norm, frequently results in the phenomenon
of Fine-Grained Forgetting. This problem is characterized by the blind discarding of filters
essential for capturing crucial local textures, culminating in the representational collapse
of the network. To mitigate this flaw, this work proposes a texture-aware pruning metric
capable of ranking convolutional filters based on their spatial activation patterns. The
methodology establishes a hybrid scoring function that integrates structural magnitude
with textural relevance, quantified by the spatial variance of latent maps through the
standard deviation. In order to avoid the destructive veto effect—a scenario in which a
low activation value would completely nullify a weight of high topological relevance—an
Additive Linear Fusion metric was developed. Additionally, to mitigate the surgical shock
of weight removal, the approach is orchestrated both in a single step (One-Shot) and in a
paced transition divided into two stages (Two-Steps). Empirical results demonstrate that
the combination of the hybrid metric with the smoothed cut stabilizes residual blocks and
prevents accuracy collapse even under severe sparsity regimes, achieving a parametric
reduction of up to eighty percent. The solution enables the generation of ultra-efficient

models with low computational cost, fully suitable for deployment on Edge Artificial
Intelligence (Edge AI) devices.

COMMITTEE MEMBERS:
Interna - 1350250 - ANNE MAGALY DE PAULA CANUTO
Interno - 2177445 - BRUNO MOTTA DE CARVALHO
Interno - 1702873 - MATHEUS DA SILVA MENEZES