Table 3 Comparison of advantages and disadvantages of each sub-method of classification network
From: Plant diseases and pests detection based on deep learning: a review
Method | Advantages | Disadvantages |
|---|---|---|
Using network as feature extractor | Obtaining effective lesion features | Relying on other classifiers for final classification results |
Original image classification | Classic in structure, it is also the basis of other classification network sub-methods and can refer to many existing networks | Lesions need to account for a certain proportion in the image, otherwise their characteristics are easily pooled out, and generally only one class of lesion is allowed in an image |
Classification after locating ROI | Obtaining ROI information of the lesions | Additional methods are needed to obtain ROI |
Multi-category classification | Solving sample imbalance to some extent | Secondary training is needed |
Sliding window | Get rough localization of lesions in images | Sliding window size requires accurate selection, and can only get rough position, slow speed of traversal and sliding |
Heatmap | Generate more accurate lesion areas | Accurate lesions location depends on network classification performance |
Multi-task learning network | Combining other networks to obtain exact location and category of lesions simultaneously, and reducing the number of training samples required | The network structure is relatively complex, and a pixel-by-pixel label is required when adding segmentation branches |