Improving CNN-Based Pest Recognition with a Post-Hoc Explanation of XAI

Abstract Deep learning is currently quite prevalent and is often used in image classification or object detection. This article adds emerging research on the use of explainable AI (XAI) in Tessaratoma papillosa pest identification and investigates the connotation and importance of XAI, interpretability classification standards, and neural network interpretation methods and compares the quality of interpretations between different approaches and various trade-offs. The experimental results include the data processing in the research, the establishment of training models, a comparison of the results and feature visualization methods, and the consequences of improving the training models. First, we analyzed the data processing methods of the dataset, trained the VGG16 model, and finally added a visual interpretation method to the model to visualize and explain the model identification results. The experimental results indicated that the best visual discrimination effect was obtained through eXplanation with Ranked Area Integrals (XRAI). In this study, XAI was used to obtain the factors contributing to incorrect predictions based on post-hoc explanations. Based on the inferenced result, we proposed an adjustment method for improving the model accuracy as a basis for future research to subsequently adjust and improve the model. It is hoped that the experimental results of this study can provide researchers in artificial intelligence useful information so that they can use XAI to acquire appropriate interpretations to correct recognition accuracy and drive the development of XAI.