Convolutional neural networks (CNNs) have made substantial contributions to the domain of plant disease diagnosis, attaining noteworthy levels of accuracy. The primary objective of this study is to enhance the capabilities of CNNs within this particular field. In this article, we suggest a unique strategy utilizing MultiRes blocks and attention mechanisms to enhance the classification performance of CNNs for plant diseases. Our solution involves stacking four MultiRes blocks, each of which gradually increases the number of filters to prevent excessive propagation of memory requirements to deeper network nodes. In order to collect more spatial information, we also implement a residual link and a 1 \(\times \) 1 convolutional layer. After each MultiRes block, we employ a Convolutional Block Attention Module to emphasize vital information and decrease redundant noise by inferring attention mappings along the channel and spatial dimensions. Finally, our method includes the use of a Dilated Spatial Pyramid Pooling module which is designed to extract information from the input image at various scales. Our evaluation leveraged a comprehensive dataset compiled from multiple sources, including the New Plant Diseases Dataset, Corn Dataset, and Coffee Dataset, encompassing a wide range of plant leaves both healthy and diseased, across various classes. This diverse collection of images, augmented by several techniques to enhance data variance and model robustness, provided a solid foundation for assessing the performance of our proposed model. Experiments demonstrate that this method outperforms conventional machine learning techniques and delivers high classification accuracy for plant diseases, with a precision of 99.34%, recall of 99.25%, f1-score of 99.29%, and accuracy of 99.35%. Our proposed model has the potential to considerably enhance the effectiveness and efficiency of diagnosing plant diseases in the agricultural industry.
