Computer Vision-Enabled Inventory Management System: A Cloud-Native Solution for Retail Cost Reduction

The computer vision data acquisition subsystem uses a distributed camera network deployed in concert with GPU edge computing modules to accomplish highly concurrent SKU image stream access and real-time acquisition preprocessing. The image input of the system is sampled at 1920×1080 resolution per frame with a sampling frequency of 2fps, and the daily processing capacity of a single node exceeds 38,000 frames.The YOLOv7-B6 model structure is used as the detection backbone, and combined with the attention-weighted channel enhancement mechanism to optimize the recognition stability in SKU-dense stacking scenarios [2]. The images are acquired, preprocessed by the embedded Jetson Xavier module and uniformly encapsulated into a multi-task tensor representation defined as: where $F_{i,j}^{sku}$ is the SKU feature response at the (i, j) location, $E_{i,j}^{bg}$ is the background interference term, $M_{i,j}^{\mathrm{meta}}$ is the image metadata, and $\alpha , \beta , \gamma$ is the weighting factor [3]. The preprocessing tensor is uniformly uploaded to AWS S3 Glacier deep-cooled storage, and the image redundancy compression ratios are listed in Table 1. The edge node state distribution and processing efficiency are shown in Figure 1. This module is designed to support asynchronous decoupling through AWS Lambda and Amazon EventBridge, ensuring data consistency and latency tolerance of the back-end inventory scheduling module in high I/O scenarios.

The cloud-native inventory management system is designed based on containerized microservice architecture, with core services deployed in AWS EKS clusters, adopting an elastic node auto-scaling mechanism, and supporting asynchronous loading of heterogeneous data structures at the SKU level [4]. The system maps the image index data $I_{m,n}$ and the dynamic inventory request vector $Q_t$ to the resource scheduling function: where $I_{m,n}$ denotes the index density of image block at m×n location, $Q_{t,m,n}$ is the request intensity at_t time, lat_t is the average response delay, and δ,μ,λ is the system scheduling weight parameter [5]. The system introduces parallel compression channels, utilizing the WebP and Zstandard bi-layer coding mechanism, which is specifically designed for redundancy-avoiding compression of SKU image data, thereby enhancing storage efficiency distribution is listed in Table 2. The data uploaded by edge nodes is unified into S3 Glacier and combined with Tagging policy to realize lifecycle management, controlling the single SKU data version to be no more than 3. AWS App Runner is used for deploying multi-tenant service routers, supporting Amazon API Gateway unified access and AWS Global Accelerator for multi-region replica consistency scheduling, and the deployment topology is shown in Figure 2.

However, the use of fixed compression parameters limits adaptability to different image structures, resulting in a suboptimal average compression ratio of 0.264 for high-resolution images. To address this, future implementations will consider content-aware dynamic encoding strategies that adjust compression intensity based on local entropy and structural redundancy. The inventory mapping engine further introduces a SKU semantic mapping building module that performs feature aggregation for each node $v_i\in V$ in G = (V,E) using a graph convolutional network: where$h_i^{\left(l\right)}$ is the features of the node_i in the first_l layer, W^(l) is the learnable weight matrix, d_i is the degree of the node i, σ is the activation function, and Figure 3 illustrates the input-output structure of the GCN module in inventory semantic aggregation [6].

Retail cost reduction is achieved through multi-source image structure compression, asynchronous data upload scheduling, and elastic service scaling, forming an integrated optimization framework, asynchronous upload scheduling and elastic scaling of service deployment [7]. The system defines the total cost function as: where $C_{\mathrm{stor}}$ denotes the storage overhead, $C_{\mathrm{net}}$ is the cost of transmitting traffic across regions, $C_{\mathrm{comp}}$ is the image encoding computational resource consumption, and $C_{\mathrm{sched}}$ is the system load cost introduced by container scheduling. $C_{xfer}$ Represents the inter-region transmission overhead caused by the migration of hot/cold data between different regions $C_{meta}$ Indicates metadata management costs generated by version control. Where $C_{\mathrm{stor}}$ is further denoted as: where $s_i$ is the compressed image volume (MB) of the first i SKU, ${\rho }_i$ is the unit storage cost (MB-day), and ${\eta }_i$ is the SKU lifecycle duration (days). The AWS Lambda-based deployment scheduling policy adopts a bandwidth-weighted minimum load prioritization mechanism, effectively balancing resource allocation in a multi-tenant environment: where $P_j$ is the prioritized scheduling probability of the service node_j, $w_j$ is the node bandwidth weighting factor, $L_j$ is the current load index, and the scheduling decision is determined by max(P_j) [8]. Compared to traditional Round-Robin and Kubernetes bin-packing schedulers, this strategy dynamically integrates real-time bandwidth constraints and load balancing, achieving up to 13.2% reduction in queue latency under 50-concurrent SKU image requests. Its adaptability makes it well-suited for heterogeneous resource environments in elastic retail deployments. The image compression part combines Zstandard entropy coding with image block level redundancy identification and the redundant region is defined as: Where ${\mathcal{B}}_{i,j}$ is the _{(i, j)} location image block, $\overline{B}$ is the block mean, $\in$ is the similarity threshold, and II is the metrics function. AWS EKS dynamic cold archive trigger controls the frequency of redundant data transfers, ensures that the compressed write rate is maintained at R_w < 18 MB/s, and uniformly controls the maximum number of SKU versions under multi-tenant conditions at K_{max ≤ 3}. The system builds a link-wide cost control strategy through function-level metrics modeling to improve cost reduction in heterogeneous regional deployment environments [9].

The experimental environment is constructed in the AWS Oregon and Singapore regions, using a dual-zone heterogeneous deployment scheme, with the server-side configuration of 32-core vCPU, 256GB memory, NVIDIA A100 GPU × 2, network bandwidth set to 1.2Gbps, and edge-side nodes deployed in parallel using Jetson Xavier AGX. The test scheme is designed as follows: (1)Build SKU image datasets with 120 categories and a total of 110,000 images, sourced primarily from Concrete-Crack500, SDNET2018, and UAV-collected images from retail store environments, to evaluate image acquisition latency and throughput; (2) test the WebP+Zstd co-coding computation consumption and storage write rate under different compression ratios; (3)perform multi-tenant dynamic concurrency simulation of the AWS SageMaker endpoint-based service scheduler to validate the container cold-start latency and request queue fluctuation response; (4) applying 50 concurrent HTTP traffic pressure for 5 minutes to monitor the average service latency and the triggering logic of cross-region load balancing policy.

The system function validation covers four major modules: image recognition, data compression, service scheduling and multi-tenant request processing. The image recognition module completes frame-by-frame detection with a frame rate of 2fps in SKU dense stacking scenarios, with an average processing time of 38ms for a single frame, and supports a number of parallel channels of up to 16; the data compression module realizes a Zstd compression rate of 12.5MB/s for a 1920×1080 image, with a block-level redundancy rejection rate of 29%~37%. The block-level redundancy rejection rate is maintained between 29% and 37%, and the average compression ratio is 0.264; the service scheduler maintains an average response latency of 85ms under the dynamic access of 50 concurrent tenants, and the service scheduler maintains an average response latency of 85ms under dynamic access from 50 concurrent tenants, with a maximum cold start delay of 280ms. This bandwidth-weighted minimum load scheduling strategy demonstrates a degree of effectiveness in resource balancing, but has not yet been benchmarked against native Kubernetes algorithms such as LeastRequestedPriority. In real-world retail scenarios—such as bulk SKU validation before promotional events or high-concurrency inventory audits—the 280ms delay remains acceptable for non-real-time workloads. However, its impact on latency-sensitive business processes still requires further evaluation through scenario-specific traffic simulations and threshold fine-tuning. the container horizontal expansion strategy triggers automatic deployment after the load factor exceeds the 0.7 threshold, and the new replica startup time is less than 3.2 seconds, and the load balance ratio is controlled within 1.08, which meets the requirement of the load balance ratio. The load balance ratio is controlled within 1.08, which meets the system-level all-link cooperative scheduling capability verification requirements [10].

The system conducts quantitative comparison experiments on image storage cost, service compute resource occupation and scheduling load in the retail cost reduction path. The average compression ratio after image compression is 0.264, and the percentage of redundant blocks of single SKU data is reduced from 38.6% to 12.4%; the average monthly write traffic of off-site cold archiving is controlled within 71.3GB. Deployment resource utilization efficiency improves to 68.5% after enabling Cloud Run dynamic replica scheduling, scheduling load balancing ratio is optimized from 1.52 to 1.08, and concurrent processing capacity under the same resource configuration is supported to increase by 39.2%. Specific indicators are shown in Table 3:

The data in Table 3 shows that the compression rate is greatly improved by the combination of image block-level redundancy detection and double-layer compression structure, while the write traffic is suppressed by the redundant partitioning strategy and delayed upload trigger mechanism. Service-side cost optimization focuses on the dynamic expansion of replica scheduling and multi-tenant scheduling deployment, so that the scheduling system can avoid the waste of peak resources under load impact, and the CPU utilization rate is effectively controlled. The overall analysis shows that the system has completed the compressive reconfiguration of the cross-domain resource structure under the premise of maintaining the service responsiveness, and supports the high-frequency image recognition retail business environment oriented to the SKU dimension.