Efficient Differentiation of Biodegradable and Non-Biodegradable Municipal Waste Using a Novel MobileYOLO Algorithm

The rapid urbanization and subsequent rise in the urban populace have thrust the challenge of the “urban trash siege” to the forefront of environmental concerns in cities. This growth of urban sectors has been linked to a staggering increase in the generation of municipal waste, intensifying the urgency of the issue known as the “urban trash siege”. Research by International Lianhe Zaobao suggests that by 2050, there will be a 70% growth in global garbage volumes [1-3]. In such a context, strategies to achieve the reduction, recycling, and harmless disposal of urban waste have emerged as imperative challenges for urban governance.

Industrialized nations have identified waste reuse as a pivotal measure in this combat [4]. The success of waste reuse is contingent on the efficient classification of trash. It has been observed that hazardous materials such as waste fluorescent bulbs, discarded thermometers, obsolete pharmaceuticals, packaging, and waste image sheets are typically consigned to landfills. Concurrently, valuable recyclable commodities, encompassing materials like paper, plastic bags, and beverage bottles, are found amidst this refuse [5]. When valuable waste is discerned and segregated from non-recyclables, it not only ensures environmental protection but also catalyzes socio-economic progress. Illustratively, one tonne of waste paper has been converted into 850 kilograms of quality paper, saving 300 kilograms of wood in the process; similarly, a tonne of discarded plastic bottles has been transformed into 600 kilograms of unleaded gasoline or diesel [6].

Waste classification undoubtedly benefits societal structures, economies, and the environment. It has been recognized for conserving natural resources and bolstering the recycling movement, while concurrently curtailing pollution [7]. Despite its significance, the predominant method of waste identification has been reliant on human efforts. Manual sorting, though widely practiced, is not only marked by inefficiencies but also poses potential health threats to the workforce [8, 9]. Given these challenges, a shift towards leveraging artificial intelligence for waste categorization has been advocated. The future of waste management is envisaged to be dominated by intelligent categorization and recycling.

Over recent years, advancements in Deep Learning (DL) have permeated a diverse spectrum of sectors [10]. Technologies underpinned by DL have been extensively deployed to address multifarious real-world challenges, enhancing system precision and efficacy in the process [11]. Such methodologies find applications across an array of sectors spanning healthcare, security, and automation, to name a few.

The automatic detection of littered areas has been explored to bolster cleaning efficiency and preclude secondary pollution. Hyperspectral data from aircrafts has been utilized to track waste dispersal across expansive terrains [12]. An unsupervised area proposal generation technique integrating selective search and non-maximum suppression has been applied to pinpoint the extent and position of littered zones [13]. Drawing from remote sensing imagery, models to autonomously discern areas riddled with waste have been developed. Preliminary experiments have demonstrated that the aforementioned model can adeptly identify littered regions with swiftness and precision, underscoring the efficiency and feasibility of deploying DL tools for such purposes.

For the real-time identification of waste, a 23-layer CNN model was employed, coupled with the generation of area proposals and the adaptation of the Visual Geometry Group-16 model for discerning plastic bottles amid challenging backgrounds [14, 15]. SpotGarbage, an innovative smartphone application, was introduced to detect and loosely classify litter locations within geotagged images captured by users. Deep convolutional neural networks (DCNNs) have been harnessed for the detection and pinpointing of pavement flaws and waste. The employed waste classification technique is anchored on the ResNet-34 algorithm [16]. Enhancements to this algorithm included multi-feature fusion of input images, feature reuse within the residual unit, and the introduction of a novel activation function [17]. The refined trash classification technique demonstrated capabilities in efficiently and accurately sorting waste.

From an environmental perspective, the repercussions of inadequate waste management are profound, manifesting as pollution, soil and water source contamination, and heightened greenhouse gas emissions. On an economic front, suboptimal waste management practices can impose considerable financial strains on municipalities. Absence of meticulous identification and segregation during sorting and disposal can inflate associated costs, encompassing transportation, landfill utilization, and waste processing. Moreover, public health is invariably compromised with substandard waste management, evidenced by disease proliferation, contamination of food and water resources, and the surge of pests.

Utilizing the KITTI dataset, an accuracy of 95.7% was achieved by the proposed system during trials. When juxtaposed with YOLOv4, the MobileYOLO model manifested significant enhancements. A notable reduction in model parameters by 53.12M was observed, amplifying computational efficiency and tempering storage and processing demands. Furthermore, a marked diminution, approximately one-fifth, in connectivity size was witnessed, fostering efficient information dissemination within the system. This model also showcased an 85% surge in speed detection prowess, catalyzing swifter, and more precise object detection and tracking. Such strides in accuracy, parameter reduction, connectivity size, and speed detection underscore the potential of the proposed system, earmarking its applicability across varied object detection realms.

Subsequent sections delve deeper into this research. Section 2 elucidates works pertinent to this domain. An innovative algorithm for the streamlined identification of biodegradable and non-biodegradable entities within municipal waste is delineated in Section 3, with results elucidated in Section 4. The manuscript culminates in Section 5, where future avenues of research are contemplated.

3.1 Lightweight model for pavement garbage classification

A model, predicated on YOLOv5, which demonstrates proficiency in rapid object detection, is introduced in this article. To tailor this model for waste categorization, structural modifications are undertaken. The ghost module is incorporated into the backbone network to curtail the model’s dimensions [27]. In an endeavor to further streamline the model, regular convolution layers are supplanted by depth-wise separable convolutional layers, leading to a reduction in the number of variables within the PANet network. To maintain detection accuracy while concurrently achieving a diminutive model size, the attention mechanism termed SELayer is assimilated into the backbone network. As illustrated in Figure 1, image features are initially extracted from the backbone network and subsequently channeled into the PANet network. In the concluding stages, the input feature maps undergo processes of up-sampling and down-sampling, ultimately yielding predictions for both category and bounding box. Notably, within the structure delineated in Figure 1, the ghost module is substituted for the BottleneckCSP, with ‘N’ signifying the number of times the module has been stacked.

1.png

Figure 1. Architecture overview

3.2 Embedding of the ghost module

2.png

Figure 2. Schematic of the ghost module

3.png

Figure 3. Depiction of the depth-wise separable convolutional model

The GhostNet, a pioneering lightweight neural network proposed by Huawei Noah’s Ark Laboratory, is predicated on the integration of the ghost module [28]. When integrated into a convolutional neural network, the ghost module has been observed to diminish computational costs and model parameters. Such reductions arise because the ghost module is capable of constructing additional feature maps via highly economical operations. For addressing memory and performance concerns associated with the garbage classification model, the BottleneckCSP component of YOLOv5 was replaced with the ghost module having step sizes of 1 and 2, respectively. As illustrated in Figure 2, feature maps within the ghost module are manipulated to generate H′W′n feature maps, post which a linear mapping of the n channel feature maps is conducted. Two ghost modules are stacked to form the Ghost Bottleneck. The initial ghost module aids in amplifying the feature map channels, while the subsequent ghost module reduces the channel count to align with the shortest pathway.

In object detection methodologies, the Ghost Module stands out as a significant technique that curtails computational costs and memory demands. This module introduces what are termed “ghost channels”, partitioning the input channels into two distinct subsets: primary and ghost. While the primary set, having a reduced number of channels, is directly linked to the subsequent layers, the ghost set independently assimilates supplementary data. Through a linear combination, these ghost channels are integrated with the primary channels, bolstering parameter efficiency. Despite its reduced complexity, this streamlined module retains its performance, rendering it apt for real-time object detection, especially in devices with stringent resource constraints. Within convolutional neural networks, the efficacy of the Ghost Module is exemplified as it provides an optimized approach to object detection tasks. In the research outlined, the depth-wise separable CNN takes precedence over conventional convolution layers, further reducing the parameters within the PANet network. Consequently, the described model’s lightweight nature is accentuated. Figure 3 illustrates the mechanics of depth-wise separable convolution, which can be discerned in two phases: depth-wise convolution and pointwise convolution.

$\left.P_{\text {conv }}=(I-i+1) \times(W-i+1) \times N_c \times i \times i \times K\right.$           (1)

$P_{K W \text { conv }}=\left((l-i+1) \times(W-i+1) \times N_c \times i \times i \times K+N_c \times 1 \times 1 \times K \times(I-i+1) \times(W-i+1)\right)$             (2)

Computational costs associated with the Depth-wise Separable Convolutional Model are exemplified in Eq. (1) and Eq. (2):

Eq. (1):

· $P_{\text {conv }}$ denotes the computational burden of the conventional convolutional procedure.

· I and W symbolize the spatial metrics of the input feature map (width and height, respectively).

· $N_c$ is representative of the input channel count.

· i signifies the kernel size (dimensions of the convolutional kernel).

· K stands for the output channel tally.

Eq. (2):

· $P_{K W \text { conv }}$ embodies the computational expenditure of the Depth-wise Separable Convolutional operation.

· The inaugural term signifies the computational load of the depth-wise convolution.

· The subsequent term conveys the computational heft of the point-wise convolution.

Subsequently, the ratio $P_{K W \text { conv }} / P_{\text {conv }}$ is computed, providing insight into the comparative computational loads.

$\frac{P_{K W \text { conv }}}{P_{\text {conv }}}=\frac{1}{N_c}+\frac{1}{i^2}$        (3)

When executing depth-wise separable convolution, an elevated count of output channels is observed. This translates to the assertion that in Eq. (3), the NC count markedly overshadows h. Consequently, the training duration for standard 2D convolution is $h^2$ times that of the depth-wise separable convolution.

3.3 Network architecture and comparative models

Conventional object detection networks, exemplified by YOLOv4, typically operate in a single step. Such frameworks enable the concurrent prediction of item categories and bounding boxes, culminating in marked enhancements in detection timeframes. The architectural representation of YOLOv4 is depicted in Figure 4.

4.png

The architecture of YOLOv4 is strategically compartmentalized into three pivotal sections: the backbone networks, neck networks, and head networks. Serving as the foundation or backbone network is the CSPDarknet53 feature extraction mechanism, composed of a quintet of modules ranging from CSP1 to CSP5. Each of these modules is constructed by alternating the stacking of CSPx blocks with CBM in lieu of CBL blocks [29]. The character “x” in CSPx denotes the number of CSP blocks. Initially, the CSP divides the input into a bifurcated branch structure. One of these branches undergoes a CBM operation, subsequently traverses the residual structure, and again partakes in a CBM operation. Post this sequence, a concatenation process is employed, resulting in the output of both branches. Within this context, CBM represents the sequence of convolution, Batch Normalization, and Mish activation functions. Conversely, CBL signifies convolution, BN, and Leaky ReLU activation functions. The CSPnet topology is integrated into the YOLOv4’s backbone network. A significant residual side, enhancing learning capacity, emerges from the counteraction of CBM with the Res unit stack [30]. Through the employment of the Mish function, which possesses neither upper nor lower bounds and is characterized by non-monotonicity and infinite-order continuity, the model is facilitated in executing regularisation and upholding gradient stability in the network. The subsequent equation delineates the loss function:

$\operatorname{Loss}=\operatorname{Loss}(\operatorname{coord})+\operatorname{Loss}(\operatorname{cls})+\operatorname{Loss}(\operatorname{conf})$         (4)

$\operatorname{Loss}(\operatorname{coord})=\lambda_{\text {coord }} \sum_{x=0}^{H \times H} \sum_{y=0}^h X_{x y}^{o b j}\left(2-w_x \times i_x\right)$          (5)

$\operatorname{Loss}(c l s)=-\sum_{x=0}^{Z l \times \| H} X_{x y}^{o b j} \sum_{c \epsilon \text { classes }}\left\{\hat{p}_x(c) \lg p_x(c)+\left(1-\hat{p}_x(c)\right) \lg \left(1-p_x c\right)\right\}$             (6)

$\begin{gathered}\operatorname{Loss}(\operatorname{conf})==-\sum_{x=0}^{H \times H} X_{x y}^{n o o b j} \sum_{y=0}^M X_{x y}^{0 b j}\left\{\hat{C}_x \lg C_x+\left(1-\hat{C}_x\right) \lg (1-\right. \\ \left.\left.C_x\right)\right\}-\lambda_{\text {noobj }} \sum_{x=0}^{H \times H \times} \sum_{y=0}^{\mathcal{K}} X_{x y}^{\text {naobj }}\left\{\hat{C}_x \lg C_x+\left(1-\hat{C}_x\right) \lg \left(1-C_x\right)\right\}\end{gathered}$             (7)

where,

· H stands for grid size;

· X designates the xth feature map;

· y represents the predicted bounding box;

· w indicates width, while I defines height;

· obj and no obj signify the constants of object presence and absence, respectively.

· The predicted and actual category classifications are respectively symbolized by Cx and $\hat{C}_x$.

· The confidence pertaining to the predicted object is represented by $p_x(c)$, and $\hat{p}_x(c)$ corresponds to the confidence of the actual object.

· Penalty coefficients _coord and _noobj are similarly defined.

The SPP structure is recognized for its significant augmentation of the receptive field, thereby inheriting enhanced contextual attributes. Responsibility for transporting the P3, P4, and P5 feature layers is attributed to the PANet. Subsequent to the facilitation of a bottom-to-top data fusion enhancement, these distinct feature layers are subjected to a down-sampling process, executed in the inverse sequence of the former step. This methodology harnesses insights from low-level feature positioning while concurrently abbreviating the path of information transmission. The loss function of YOLOv4, denoted as “Loss”, amalgamates the regression loss function, probability loss function, and classification loss functions.

3.4 Architecture of the MobileYOLO model

In the present study, the MobileYOLO algorithm has been introduced, derived from the foundational structure of YOLOv4. Figure 5 provides a graphical representation of the overall architecture, highlighting the lightweight optimizations incorporated within MobileYOLO. A reduction in the model parameters of YOLOv4 was achieved by replacing its conventional communication system with the more lightweight Mobilenetv2. Furthermore, the adoption of Depthwise Separable Convolution in lieu of traditional pooling layers, both within the feature extraction network and the detection head network, contributed to this parameter minimization. To enhance the neural network, a concise attention module, termed ECANet, was integrated. This module is postulated to augment the prominence of pertinent features, attenuate redundant features, and amplify the network’s feature fusion characterization capabilities.

5.png

Figure 5. Architecture of the MobileYOLO model

3.5 Structure of the MobileNetV2 network

For the objective of extracting object features specific to waste management scenarios, MobileNetV2 has been utilized as an advanced backbone network. The adoption of this network is noted to significantly reduce the quantity of model parameters. Distinct from its predecessor, MobileNetV1, MobileNetV2 incorporates an inverted residual module furnished with linear bottleneck blocks. This inclusion is reported to markedly elevate the efficiency of image classification and detection tasks, especially when executed on mobile platforms.

The propensity for the loss of characterization trait information becomes pronounced when operations employ the ReLU activation function on layers proximate to the base convolutional layer. However, it is observed that the deployment of the Linear transfer function on high-dimensional convolution operations minimizes data loss. This phenomenon can be attributed to the richer data content inherent to the high-dimensional fully connected layer, compared to its lower-dimensional counterparts [31]. In an attempt to circumvent excessive feature information loss, the ReLU6 activation function was integrated into the high-dimensional convolution layer of MobileNetV1.

The topology of bottleneck sections within the MobileNetV2 network is illustrated in Figure 6. The primary objective of the inverted residual module is discerned to amplify the network’s feature extraction potential and facilitate efficient multi-layer information relay.

When assessing a multitude of algorithms on ImageNet-compiled data, the TOP-1 predictive accuracy serves as a measure to gauge the congruence between the primary rated classification and the factual outcome, as detailed in Table 1. Subsequent to the integration of the ECA module, a marginal increment in both the parameters and computational magnitude is detected. Nevertheless, ECA retains the potential to deliver augmented detection accuracy, all the while imposing only a trifling computational demand, thereby preserving its lightweight attributes and ensuring proficient performance.

6.png

Figure 6. Structure of the bottleneck block in MobileNetV2

Table 1. Performance indicators of various attention systems

In the quest to circumvent information loss during the dimensionality reduction process, feature information acquisition was found to obviate the need for manual cross-validation adjustments. Subsequently, the Sigmoid activation function was employed, yielding dimension c with a novel parameter. The reallocation of weights across different channel characteristics ensued, wherein non-essential features were observed to be more robustly suppressed, amplifying the significance of pertinent information. ECA, in comparisons drawn with attention modules, was discerned to mirror the efficiency and lightweight properties akin to SENet.

$H=\varphi(C)=\left|\frac{\log _2(C)}{\gamma}+\frac{L}{\gamma}\right|_{O D D}$         (8)

For multiple channels symbolized as c, the notation |*| represents the odd number proximate to the value, as depicted in Figure 7.

Within the delineated context module, convolutions of size 3*3 were leveraged, aiming to emulate the functionality of 5*5 and 7*7 convolution kernels. Such an approach was discerned to act as a surrogate for larger convolutions. Evident from Figure 8, the kernel was subsequently aimed at diminishing the aggregate parameter count. With the ISSH context module’s integration into each triad of PAnet outputs, an escalation in the overall parameter number was noted. This study saw a reduction in the SSH module output to 76*76, enhancing the feature capturing capacity for diminutive objects.

7.png

Figure 7. ECA structure

8.png

Figure 8. ISSH structure

From a managerial perspective, the deployment of this algorithm was correlated with the optimization of waste management protocols. Such implementations allowed municipalities to channel resources judiciously, predicated on the precision-driven classification of waste constituents. This precision not only translated to economic efficiency but also fostered targeted waste treatment methodologies, abating environmental ramifications and fostering sustainable modalities. Additionally, superior waste segregation facilitated by the algorithm was linked to public health enhancements, curtailing disease propagation and contamination threats. These implications are instrumental in refining waste management infrastructures, proffering multifaceted benefits to both urban centers and the environment.