A Novel One-vs-Rest Classification Framework for Mutually Supported Decisions by Independent Parallel Classifiers Antonios Vogiatzis School of ECE Technical Univ. of Crete Chania, Greece anvogiatzis@intelligence.tuc.gr Georgios Chalkiadakis School of ECE Technical Univ. of Crete Chania, Greece gehalk@intelligence.tuc.gr Konstantia Moirogiorgou School of ECE Technical Univ. of Crete Chania, Greece dina@display.tuc.gr Michalis Zervakis School of ECE Technical Univ. of Crete Chania, Greece michalis@display.tuc.gr Abstract—We put forward a generic classification architecture of independent parallel CNNs that explicitly exploits a “mutual exclusivity” or “classifiers’ mutually supported decisions” prop- erty underlying many dataset domains of interest, namely that in many cases an image in a given dataset might almost unques- tionably belong to one class only. Our framework incorporates several designed-to-purpose opinion aggregation decision rules that are triggered when the mutual exclusivity property is or is not satisfied; and makes use of “weights” which intuitively mirror the confidence each CNN has in identifying its corresponding class. Our framework can thus (a) take advantage of clear class boundaries when these exist, and (b) effectively assign items to classes with increased confidence, even when clear class boundaries do not exist. We confirm the effectiveness of our approach via experiments conducted on a well-known dataset from the waste classification domain. Index Terms—image classification, supervised learning, mutual exclusivity, decision rules, opinion aggregation I. I NTRODUCTION Many image datasets typically contain images with dras- tically different characteristics, allowing for the training of different classifiers with each of which succeeding in achieving high confidence levels. Examples include (a) recyclable materials: for instance, glass objects are very different from plastic, making it easy to differentiate between the materials; (b) facial recognition problems: e.g., the color of the skin and the shape of the eyes are powerful characteristics which facilitate the required separation; and (c) X-ray datasets: in such images, the bone and tissue density provide strong signals that can unambiguously denote certain medical conditions. In this paper we put forward a generic classification frame- work that is well-suited for such settings. Our approach casts a single label multi-class classification problem into one to be tackled via the use of as many parallel binary classifiers as the different classes. Our framework makes use of a neural networks-based architecture which tackles the multi- class classification problem as one composed by multiple This research has been co-financed by the European Union and Greek na- tional funds through the Operational Program Competitiveness, Entrepreneur- ship and Innovation, under the call RESEARCH—CREATE—INNOVATE (project code: T1EDK- 03110, ANASA). “binary” classification ones. As such, our approach belongs to the one-vs-rest family of classification methods. Importantly, our classification framework comes complete with a set of decision rules which correspond to social choice functions (or “voting rules”) [1] that enable the final “collective decision” on the choice of the single-to-be-assigned label, given the outputs of the “dedicated per class” binary classifiers. Specifically, our proposed architecture consists of a set of independent dedicated binary classifiers, each one with two output nodes, with each output node expressing the probability of the item under consideration to belong in the class. Our key intuition is that when only one of the independent dedicated classifiers, say k, puts a “large enough” probability on the item under consideration belonging to its class, while all others believe the item cannot be classified in their respective class, then we can be confident that the class to select as the output of our system is indeed the one predicted by the k-th classifier. This is because the classifiers’ output essentially mutually support eachother. We act upon this intuition via a decision rule, MSDR (for “mutually supported decisions rule”) we put forward, and which implements it. Arguably, there are many settings and problems where the assumption behind this approach is valid and helpful; several such domains were mentioned in the beginning of this paper. Of course, there exist many cases in which features might be such that it is harder to distinguish among classes. That is, either the class boundaries are unclear, or the confidence of certain classifiers is low. The latter could occur for instance in the case of highly unbalanced datasets. Our approach is generic enough to tackle such cases, via the incorporation of (i) classifier-specific “confidence-related weights”, and (ii) several alternative decision rules we propose, and which complement the main MSDR rule mentioned earlier. Given the above, our main contribution is putting forward a generic architecture that explicitly exploits a “classifiers’ mutual support” or “mutual exclusivity” property underlying many dataset domains of interest: the fact that in many cases an image in a given dataset might almost unquestionably belong to one class only. Our architecture is able to separate