An improved efficientnet-B5 for cucurbit leaf identification

Abstract

Plant diseases significantly impact the quality and productivity of crops, leading to substantial economic losses. This paper introduces two enhanced EfficientNet-B5 architectures, EfficientNetB5-sigca and EfficientNetB5- sigbi, specifically designed to detect and classify diseases in cucurbit leaves. We employ EfficientNet-B5 for feature extraction, using a 456×456×3 input and omitting the top layer to generate feature maps with Swish activation. A global average pooling 2D layer replaces the conventional fully connected layer, producing a flattened vector. This is followed by a dense layer with four output units, L2 regularization, and sigmoid activation, using either categorical or binary cross-entropy as the loss function. We also developed a novel image dataset targeting cucumber and cantaloupe leaves, including 11,425 augmented images categorized into four disease classes: anthracnose, powdery mildew, downy mildew, and fresh leaf. Our experiments dataset demonstrates that the EfficientNetB5-sigbi achieves an accuracy of 97.07%, marking a significant improvement in classifying similar diseases in cucurbit leaves.