Evaluating ResNet Network Performance: Accuracy, Loss, and Recall Explained

In order to evaluate the classification performance of the improved ResNet network, evaluation metrics need to be set. The commonly accepted evaluation metrics in the academic community are accuracy, loss, and recall. In this article, accuracy and loss are discussed as evaluation metrics along with recall.

Assuming there are 10 classification categories, and there are 10 test images, when one image is input into the network, 10 category probabilities are obtained. The 'top-5' refers to the top five categories with the highest probabilities among these 10 categories. If the category of this test image is among these five categories, it is considered a correct prediction, otherwise, it is a wrong prediction. The 'top-5' error rate is the number of wrong predictions divided by all samples, and the 'top-5' accuracy rate is the number of correct predictions divided by all samples.

The 'top-1' refers to the category with the highest probability among these 10 categories. If the category of this test image is the category with the highest probability, it is considered a correct prediction, otherwise, it is a wrong prediction. The 'top-1' error rate is the number of wrong predictions divided by all samples, and the 'top-1' accuracy rate is the number of correct predictions divided by all samples.

Recall is a measure of coverage, which measures the proportion of true positives among all actual positives. It can be seen that recall considers the proportion of true positives recalled. Its formula can be derived as follows: