Total contribution score and fuzzy entropy based two-stage selection of FC, ReLU and inverseReLU features of multiple convolution neural networks for erythrocytes detection

The proposed system aims at automatic erythrocytes detection using ensemble of selected features of multiple convolution neural networks (CNNs) to overcome the shortcomings of existing works arising due to the highly overlapping characteristics of handcrafted features. The main merit of this work lies in the proposed two-stage feature selection algorithm, which completely eliminates the chances of information loss inherent in traditional CNNs, occurring due to the suppression of negative values of features by rectified linear unit (ReLU) and also largely reduces the feature dimensionality. Moreover, it is the first algorithm proposed, which is capable of selectively suppressing the positive or negative values or none of each feature depending upon its respective significance in classification, in contrary to the previously proposed variants of ReLU. Firstly, it constructs a feature space for each CNN by performing inter-selection among its Fully connected, ReLU and InverseReLU features and selecting features possessing minimum Fuzzy Entropy and maximum newly formulated Total Contribution Score values simultaneously. Secondly, it performs intra-selection within each selected feature space, eliminating less significant features which simultaneously satisfy the redundancy and non-relevancy criteria stated here. Finally, by performing detection using the feature-ensemble, this method registers 98.6% mAP, proving its excellence over existing works.