Dual branch CNN with attention model using an Adaptive Multi-scale Edge Aware Feature

Abstract—Neural networks have proved their usefulness in
many applications. However, due to the noisy nature of most realworld
images, most algorithms adapt various feature extraction
and pre-processing techniques in order to achieve better results
from their model. In this work, we proposed an innovative dualbranch
Convolutional Neural Network (CNN) with spatial and
channel attention-based calibrations. The model processes high
and low-resolution inputs through parallel branches to capture
both fine-grained and contextual features effectively. Our model
also incorporates an adaptive Edge-Aware Multiscale Feature
Integration (EAMFI) used in several stages of the CNN learning
stage. This technique helps to specifically amplify diseaserelevant
signals in the images plagued by variable lighting and
background noise. The implementation of multiscale detection
produces a feature-rich input for our dual-branch, attentionenhanced
CNN to exploit. We evaluate the model on the maize
leaf images captured under diverse field conditions, representing
four distinct disease categories. The proposed architecture is
efficient in detecting and predicting maize diseases compared
with other detection models. Ablation studies confirm that the
integrated attention mechanism and EAMFI pre-processing are
critical contributors to this performance. The resulting framework
provides a scalable, interpretable solution for precision
agriculture, with a lightweight implementation enabling efficient
on-device inference in resource-constrained settings.
Index Terms—Maize Disease Classification, Attention Mechanisms,
Dual-Input CNN, Edge-Aware Preprocessing, Precision
Agriculture