<![CDATA[The Squeeze-and-Excitation (SE) Block, a concrete form of channel attention mechanisms, serves to flexibly readjust feature channels, with squeeze using Global Average Pooling for global context extraction across all channels. The limitation of the SE Block lies in discrete spatial merging, which causes important information to be lost. This research proposes an Integral-based SE Block by formulating the Squeeze operation as a continuous spatial integral to produce clearer channel feature merging. The IntSE block is implemented on the densenet121 architecture by replacing the discrete spatial SE without changing the structure of the backbone. The effectiveness of the architecture proposed in this study uses secondary datasets, namely Plant Village and HM10000, which were used for experiments during evaluation. The PlantVillage dataset consists of 12,246 tomato leaf data and consists of 10 classes, while the HAM10000 dataset consists of 10,015 data and has 7 classes. The results show that DenseNet121 with IntSE achieves an accuracy of 99.25% and an F1-score of 99.25% for the PlantVillage dataset, as well as an accuracy of 83.57% and an F1-score of 83.19% for the HAM10000 dataset. This experiment shows that densenet121 with IntSE outperforms basic densenet121, Discrete Spatial SE, and Attention CBAM in terms of accuracy and F1-Score. The ablation analysis shows that the formula of spatial squeeze integral performance is improved compared to global pooling and adaptive pooling. Statistical testing shows significant changes to state that the performance of the actual architecture.]]>
