Plastic waste presents a growing environmental challenge due to its persistence and increasing accumulation, making efficient and accurate classification of plastic types–including PET, HDPE, PVC, LDPE, PP, PS, and others– essential for real-time automated recycling systems. The inherent heterogeneity of plastic types presents significant classification challenges that a single network cannot adequately address. In this work, we propose a class-aware multi-expert knowledge distillation framework to deploy high-capacity models on resource-constrained edge devices. Accordingly, specialized models have been developed, each serving as an ’expert’ for specific plastic classes. By leveraging these ensemble classifiers within a multi-expert teacher knowledge distillation framework, we have introduced a compact student model, further optimized into an even tinier variant suitable for real-time embedded deployment. The performance of such models is heavily contingent upon the quality of the training dataset. This challenge is particularly pronounced in plastic waste due to the high morphological variability of materials and their susceptibility to deformation or alterations in appearance caused by aging and contamination. Consequently, existing public datasets often lack the diversity required for robust model generalization. To address this limitation, we have consolidated multiple datasets into a unified framework, providing a valuable benchmark for automated plastic waste classification. The generalizability of the proposed framework has been demonstrated across multiple student and teacher models using both homogeneous and heterogeneous model backbone architectures. Experimental results demonstrate that the proposed distillation approach enables the tiny student to achieve high accuracy and robustness while significantly reducing model size (by 8.3) and inference time.

Waste management is becoming an intriguing problem all over the world. waste’s volume is becoming more and more huge due to the urbanization and industrialization. This issue impacts all components of the environment, including air, water, and land. This alarming situation needs some solutions which should be rapid, urgent and innovative. For this reason, several solutions have been proposed to reduce the huge amount of waste globally. In fact, recycling may be an efficient tool, especially when the sorting process is done automatically. As this is one of the main goals of the circular economy, many researchers have focused on developing automatic solutions to address environmental problems. Recently, with rapid advancement of Artificial Intelligence (AI), different solutions were developed in order to improve accuracy and save time. Many investigations on AI techniques have been made in order to include machine learning, deep learning and ensemble approaches. This paper reviews AI based approaches for waste classification while emphasizing their real world scope. In addition, it highlights the most popular dataset used in order to validate the efficiency of the developed solutions. The main goal of this paper is to provide an overview of the existing techniques of AI helping researches to built a more accurate solution for a cleaner environment.

Waste is causing a serious environmental problem. Waste management and circular economy offers a best alternative for preserving the environment and protecting nature resources. This paper presents an AI-based system for waste identification and sorting, particularly focusing on plastic, which is the less decomposed waste, using computer vision techniques. The data, used for system validation, comprises publicly available datasets and a locally collected dataset, devoted to plastic classification and detection. Performance metrics demonstrate that YOLOv8 outperforms other deep learning solutions, in real-time waste detection and is well suited for implementation on a Raspberry Pi 4 Model B. The system outperforms also existing methods in terms of efficiency, accuracy, and speed, and is a promising solution for addressing the global waste management challenge.

Plastic waste is becoming nowadays a huge environmental problem. Circular economy seems to be the best choice for preserving the environment, by reinserting plastic waste into the industrial process and use. A key step for that is the plastic waste segregation from other waste. This paper proposes a new plastic waste segregation approach based on two modalities: spectroscopy and optical images. The research handles an optical camera for waste image acquisition, while the ground truth for plastic type identification, with respect to the seven primary plastic types: PET, HDPE, LDPE, PP, PS, PVC, and Other, is established through spectroscopic analysis. Various deep learning models were trained and evaluated on a dataset of publicly available and custom-collected images, aiming to achieve human-level accuracy in distinguishing between plastic types. The proposed methodology is based on a set of steps, where the model is first built implicitly on a segregation of different wastes, and then applied via transfer learning for separating the seven plastic types. The system performances demonstrate the feasibility and effectiveness of this combined approach for accurate and efficient plastic waste sorting. The proposed system has been implemented on Raspberry Pi 4 Model B+ as the central processing unit, integrating computer vision and deep learning techniques for efficient and accurate sorting.

Visual examination of defects in the industry is a tedious task requiring great attention and time to scrutinize several parts. That’s why implementing an automated system for defect detection should be placed to mitigate such disruption in the production line. Such systems leverage several image processing techniques in order to identify and extract defect areas. Unfortunately, the similarity between defect types makes defect classification very ambiguous. To cope with such a limitation, possibility theory offers a modeling space where uncertainty and ambiguity can be conveniently handled. In this context, the Dubois and Prade fromalism is applied on texture feature’s regions containing defaults basing on Local Binary Pattern (LBP). Accordingly, a possibility mapping of the different defects, which yields a more reliable decision making within the possibility/necessity paradigm. This approach’s effectiveness is demonstrated through a comparative study against other state-of-the-art techniques. The proposed method outperforms others by achieving a mean Precision mP of 81.6% on the NEU-DET database.

Breast cancer early diagnosis is a major concern for reducing deadly cases. Automation of microcalcification detection is becoming increasingly important given their tiny scale. At this step, a high false negative rate is observed, leading to high ambiguity. In this context, conventional approaches are unable to handle such ambiguity. Possibility theory offers a powerful paradigm enabling to handle a high uncertainty level. Therefore, in this research,new possibilistic modeling strategy for microcalcification detection is proposed. The developed system is based on k-means clustering followed by a fusion of the corresponding possibility distributions, for decision making. For enhancing classification's accuracy, two aspects may be taken into consideration: scattering within classes and the class discrimination power. A high inter-class variance may be regarded as an evidence of good discrimination. Nevertheless, in case of existing highly scattered classes, miss-classifications of samples at the class margins are still encountered, even at high inter-class variance. Clustering solves this problem by redefining classes in a more compact way. Under this paradigm, the clustering optimization is performed using a criteria based on the Area Under Curve(AUC), where the confidence degree level is based on the consistency principle of Dubois and Prade. The above strategy has been applied for the detection of microcalcifications, based on a set of pyramidal and multi-resolution based features. Validation on the Digital Database for Screening Mammography (DDSM) public dataset has been undertaken. The proposed system, which gives 99.4 % detection accuracy, can be used to assist medical practitioners.

Dendritic cells can be seen as a mirror of our immune system. Based on their in virto analysis, biological experts are now able to study the impact of food contaminants on the human immune system. Accordingly, a visual characterization of dendritic cell morphology can provide an indirect estimation of the toxicity. In this paper, we propose an automatic classification of dendritic cells that could serve as a second non-subjective opinion for pathologists. The proposed approach is built on pre-processing steps for segmentation and cell detection in microscopic images. Then, a set of features such as shape descriptors are extracted for cell characterization. At this step, three cell classes are distinctively identified by experts. Nevertheless, a high ambiguity is revealed between cell classes. Possibility theory can offer a realistic framework for making reliable decisions under high ambiguity. It exploits a human natural concept of the implicit use of probability distribution for deciding on the possibility of some assertions in some contexts where a cognitive conflict is observed while interfering existing related postulates, leading to high ambiguity. Based on the consistency concept of Dubois and Prade, a transformation of the probability into a possibility distribution is undertaken. Under possibility paradigm, a further feature selection in the possibility space using the Shapely index. Compared to state-of-the art methods the proposed approach yielded on a real dataset of nearly 630 samples an improvement in terms of the mean precision rate, the Recall rate, and the F1-measure.

The biometric systems face variability, incompleteness and insufficiency in data, which affects the performance of the recognition system. In iris recognition systems, several conditions cause different types of degradations on iris data such as the poor quality of the acquired pictures, the iris region which can be partially occluded due to light spots, or by lenses, eyeglasses, hair or eyelids, and adverse illuminations or contrasts. All of these limitations are open problems in the iris recognition and affect the performance of iris localization, iris feature extraction or decision making process, and appear as imperfections in the extracted signature. This paper addresses the use of the uncertainty theory for modeling iris system imperfections. Several comparative experiments were conducted on three subsets, namely CASIA.Ver4: synthetic, thousand and interval iris databases. Experimental results show that our proposed system, based on the possibility theory, improves the iris recognition system in terms ROC, AUC, FAR, FRR and PIN, compared to other iris identification systems.

For a safe navigation of visually impaired people, traversable area segmentation seems well important for preventing collision or falling down. In this respect, reducing the system cost entails applying low cost low resolution monocular cameras, leading to poor quality images. Adding potential navigation associated camera vibration, the system proves to be liable to some data imperfections. Such image affecting imperfections refer well to the necessity of applying the possibility theory as an appropriate means for reducing information ambiguity. In this regard, a new traversable area segmentation based on possibility modeling theory approach is being developed in this work. Noteworthy and for satisfactory adaptation of our model to image condition variability, a crucial starting point is imposed, namely considering reference area placed in the bottom of the image, assumed to be a traversable area. Actually, the proposed approach involves three major steps, namely: a color feature extraction step, a possibility distribution generation step, based on Dubois and Prade principle, as well as a data fusion step where a conjunctive operator is achieved for generating a distinctive unique possibility map. To note, the proposed approach is validated on a publicly as well as a developed databases. A mean accuracy rate of 98% is obtained with respect to both databases.