Enhancing IoT-Driven Garbage Analysis in Smart Cities Through Interpretable CNN Feature Maps

In the contemporary era, the pressing challenge for an increasingly urbanized society is the transformation of cities into technological green islands that facilitate a peaceful and sustainable lifestyle for their inhabitants. Among the myriad of issues plaguing large metropolises, the accumulation of uncollected waste stands out as a critical concern that not only detracts from the aesthetic appeal of urban environments but also poses significant health and environmental risks. Fortunately, advancements in modern computer vision systems have reached a level of sophistication that empowers them to efficiently identify and classify waste through photographs captured by surveillance cameras. This paper undertakes a comprehensive evaluation of the application of convolutional neural networks (CNNs) as a foundational element for developing a garbage detection system based on image data from streets. We delve into various architectural strategies for constructing an IoT solution that seamlessly integrates a high-resolution camera, a compact microcomputer, a pre-trained neural network, and a service for incorporation into a smart city cloud infrastructure. To elucidate the inner workings of the neural network and its interpretative processes at various layers, we employ advanced visualization techniques to generate feature maps for all object classes relevant to our study. We show that a CNN used for garbage classification acts in a bio-inspired way.