Application of Computer Image Recognition Algorithm in Agricultural Product Quality Inspection

Abstract

In agricultural product quality inspection, subtle defects of agricultural products are easily disturbed by complex background and lighting changes, making them difficult to identify. To solve this problem, this paper uses Vision Transformer (ViT) to increase the precision of identifying minute flaws on agricultural items’ surfaces. The training data set is enlarged using data augmentation technology, which rotates, crops, and modifies brightness to enhance the model’s capacity to adjust to various changes. Using a self-attention mechanism, the ViT model detects minor surface flaws globally and captures long-range dependencies in the image. Combining transfer learning and fine-tuning strategies, the ViT model pre-trained on a large-scale image dataset was optimized on a specific agricultural product dataset, further improving recognition accuracy and robustness. Experimental results indicate that the ViT model in this work has F1-score of 0.89 and a precision of 92.5% when compared to other models. This shows that ViT has high accuracy and robustness in agricultural product quality detection. The results of this paper have made a great contribution to the development of future agricultural automation detection.