Transfer Learning Garbage Classification System: MobileNetV3 Production Design with Dataset & Metric Specs

System Overview

This article presents a complete garbage image classification system built on transfer learning. The system includes a user-facing frontend for image upload, recognition, correction, and personal management, as well as an admin backend for user management, category configuration, knowledge base, dataset and model lifecycle management, and system monitoring. The technical route uses ImageNet pre-trained weights fine-tuned on a garbage classification dataset, with MobileNetV3 recommended for online inference and ResNet50/EfficientNet B0 as experimental baselines.

System Demonstration

Frontend User Features

Authentication

Username/password login; username ≥ 3 characters, password ≥ 6 characters.

User registration assigns ordinary role; admin accounts cannot be registered on frontend.

Successful login saves token and redirects to source page; expired token redirects to login.

Logout requires secondary confirmation, then clears token and returns to home.

Intelligent Recognition

Drag-and-drop or click to select local images; no camera/photo entry allowed.

Pre-upload validation of format, MIME type, file size, and image parsability; clear error messages with retry option.

UI shows upload guidelines (clear image, single subject, three suggestions) and the upload → AI classification → suggestion flow.

After selection, displays thumbnail, filename, dimensions, and cancel option; shows "recognizing" state on submit.

Recognition Results and Correction

Result includes original image, predicted main category, fine-grained category, confidence, model version, recognition time, and disposal suggestion.

Shows Top 3 candidate results; confidence below 65% triggers warning state and prompts user to follow local rules.

User can confirm correctness or submit correction by selecting correct category, adding explanation, and optionally uploading supplementary images.

Correction creates a feedback record with initial status "pending".

Classification Guide and Details

Keyword search (e.g., milk tea cup, used battery, courier box); shows not-found prompt and feedback entry when no results.

Four main category cards (recyclable, hazardous, kitchen waste, other) are clickable entry points to detail pages.

Detail page includes category definition, common items, disposal tips, common misconceptions, and a favorite button.

Recognition Records

Users view only their own records, filterable by date range, category, and item name.

Clicking a record opens a side drawer with original image, category, confidence, suggestion, model version, and feedback status.

Deletion requires confirmation and only removes the current user's visible record.

Personal Center

Profile: edit avatar, nickname, email, phone; registration time read-only.

Favorites: search/filter saved items, unfavorite, and navigate to disposal suggestions.

My Feedback: view pending, accepted, rejected feedback and admin handling results.

Account Security: change password, logout; irreversible actions require secondary confirmation.

Admin Backend Features

Dashboard : shows registered users, today's recognitions, total recognitions, category distribution, low-confidence count, pending feedback, current model version.

User Management : paginated query, search, usage statistics, enable/disable accounts, reset passwords; no plaintext passwords displayed.

Category Management : maintain four main categories and fine-grained categories with name, color, icon, description, sort order, enabled status, and mapping to main categories.

Knowledge Base Management : maintain item names, keywords, example images, classification descriptions, disposal suggestions, common misconceptions, and publish status.

Recognition Record Management : filter by user, time, category, confidence; view original image, result, model version, user feedback.

Feedback Audit : handle correction feedback, mark as accepted, rejected, or needs supplement; accepted samples can enter candidate training set.

Dataset Management : upload and audit samples, maintain labels, source, dataset split, sample statistics, and dataset version.

Model Management : maintain model name, version, weight location, training set version, metrics, enable/disable status, and online model switching.

System Configuration : configure image upload limits, low-confidence threshold, announcements, and local classification rule descriptions.

Operation Logs : record admin login, deletions, audits, model switches, configuration changes, and other key operations.

Business Process and State Rules

Image Recognition Process

Upload : validate local image format, size, and parsability → output success or failure reason.

Preprocessing : compress, correct orientation, resize to model input size, normalize → output inference-ready image tensor.

Model Inference : call currently enabled garbage classification model, compute class probabilities → output Top K categories and probabilities.

Rule Mapping : map fine-grained category to four main categories, associate disposal suggestions and knowledge entries → output main category, fine category, suggestions.

Record : save user, image, prediction, confidence, model version, timestamp → output recognition record ID.

Feedback : user confirms or corrects; admin audits corrections → output feedback status and candidate training samples.

State Definitions

Recognition Task : pending upload, validation failed, recognizing, success, failure. UI must clearly show current state; failure allows retry.

User Feedback : pending, accepted, rejected, needs supplement. Only admin can change processing result.

Dataset Sample : pending review, trainable, rejected. Only trainable samples can enter training data version.

Model Version : draft, evaluated, enabled, disabled, archived. Only one online enabled model allowed at any time.

Model and Dataset Requirements

Model Technical Route

The system adopts supervised image classification with transfer learning. It uses ImageNet pre-trained weights as initial parameters and fine-tunes on a garbage classification dataset. MobileNetV3 is recommended as the primary online inference model for its balance of speed and accuracy, while ResNet50 and EfficientNet B0 serve as experimental baselines for comparison.

Model Comparison

MobileNetV3 : Lightweight classification network; used as the main model for user-facing online recognition.

ResNet50 : Classic deep convolutional network; used as experimental baseline for paper comparisons.

EfficientNet B0 : Balanced accuracy and parameter count; optional enhanced model for comparing transfer learning approaches.

Dataset Requirements

TrashNet can serve as a material classification baseline, but its six material labels do not directly map to the domestic four-category waste system.

The system must supplement or build local samples for kitchen waste, hazardous waste, etc., and maintain mapping from fine-grained categories to the four main categories.

Training, validation, and test sets should be split 7:2:1 or 8:1:1; similar images must not leak across splits.

Each sample must record image path, fine-grained label, main category label, source, audit status, and dataset split.

Quality Metrics

Four-category Accuracy : Test set accuracy ≥ 85%.

Recyclables Accuracy : Common recyclables recognition accuracy ≥ 88%.

Evaluation Completeness : Output precision, recall, F1 score, confusion matrix, and per-class metrics.

Inference Latency : Average response time per image ≤ 3 seconds in a typical deployment environment.

Rejection Prompt : Results below the confidence threshold must display an uncertainty warning and candidate categories.