The success of deep learning methods has led to the development of several applications for the automated identification of plant diseases and pest attacks. Nevertheless, these programs frequently experience overfitting, and when applied to test datasets from unfamiliar contexts, the diagnostic performance is significantly reduced. In this work, we present traditional CycleGAN, a unique image-to-image translation system with an attention mechanism of its own. CycleGAN is a data augmentation technology that improves the effectiveness of plant pest diagnosis by transforming a limited number of pest-damaged images into a broad variety of pest-lesioned images. In our work, CycleGAN outperformed other models in generating synthetic images of Sawfly pests. On the other hand, the copy-paste-blend (CPB) approach has proven effective in seamlessly embedding pest masks into external leaf images. This technique blends pest masks at various scales with leaf backgrounds, resulting in synthetic images that appear more natural and realistic.