Smart waste classification and image retrieval are important for recycling, resource recovery, and circular-economy applications. This study presents a comparative deep learning framework for waste classification and content-based image retrieval (CBIR) under geometric transformations. Rather than proposing a single fused architecture, the study systematically investigates three key design factors: backbone architecture, classifier-head design, and loss-function selection. A custom waste image dataset comprising 1000 images across 10 waste categories was used throughout the experiments. The evaluated backbone architectures included ResNet18, ResNet34, and ResNet50, while the classifier heads consisted of a fully connected classifier, a DenseNet201-based classifier, and a VGG16-based classifier. In addition, three loss functions—cross-entropy, label smoothing, and ArcFace—were evaluated under identical experimental conditions. The results show that ResNet50 combined with a DenseNet201-based classifier and ArcFace achieved the highest classification performance, reaching a 98.4% accuracy and 98.3% F1-score. ArcFace also improved feature discriminability and enhanced retrieval effectiveness, with the best configuration achieving a 98.1% mean average precision (mAP). Robustness evaluation under rotation and scaling transformations demonstrated that the proposed framework maintained stable performance under geometric variations. Although ResNet50 achieved the highest accuracy, ResNet34 combined with a fully connected classifier and ArcFace provided a favorable balance between recognition performance and model complexity. These findings demonstrate that backbone depth, classifier-head design, and loss-function selection all play important roles in determining waste classification and retrieval performance.
27 July 2025 in ECTI Transactions on Computer and Information Technology (ECTI-CIT)
Computer vision is crucial in identifying and diagnosing diseases like skin cancer, which can become life-threatening if not detected early. Although numerous methods have been developed, these techniques often face challenges due to the varied nature of skin cancer, which frequently presents irregular shapes and ambiguous structures. In this study, we introduce the design of an unsupervised two-stream network capable of simultaneously learning from datasets of various sizes. The network parameters are organized from smallest to largest to improve the efficiency of feature extraction. Additionally, the network incorporates residual blocks, bidi- rectional long short-term memory, and an attention layer to help reduce training loss. The proposed method was tested using the PAD-UFES-20 dataset, using mean square error to measure training loss and accuracy to check how well it recognizes skin cancer. The results showed a loss of 0.0079 and a training time of only 0.53 minutes, performing better than other advanced methods in both loss and speed. Our approach showed better results than previous methods, accurately recognizing skin cancer and showing potential for use in a mobile app to help with early detection and diagnosis.