| The task of agricultural image classification has always been a popular topic in agricultural research. Both traditional and deep learning-based methods have emerged to address this task. However, as these methods have expanded, they have become more reliant on data and require additional external information to improve performance. In reality, agricultural images often have low quality and lack annotations, and it is challenging to obtain clear external prior knowledge and semantic information. Therefore, we aim to improve image classification using only the simplest agricultural image dataset, which consists of images and their corresponding class labels. By leveraging the information inherent in the images themselves, we seek to obtain prior knowledge and semantic information to enhance image classification performance, and we use Class Activation Mapping to illustrate the results and the improvement. Furthermore, we enhance the feature extraction process by utilizing it. We conducted experiments on four agriculture-related datasets, using Residual Neural Network as our baseline. The results show that our method achieves improvements in both Top-1 accuracy and Mean Average Precision metrics. |
*** Title, author list and abstract as seen in the Camera-Ready version of the paper that was provided to Conference Committee. Small changes that may have occurred during processing by Springer may not appear in this window.