PDF Page Splitting and Target Range OCR Extraction: Resource Optimization

1. The Memory Threshold of Browser-Based OCR

Local document scanning relies on client-side memory. When you open a PDF, the PDF.js parser reads the file stream, extracting page layers. To run OCR, the browser must render each page to a separate canvas element. This operation requires considerable memory space, particularly when processing high-resolution documents.

For a 50-page document, rendering canvases at high resolution (required for accurate character matches) can consume several gigabytes of RAM. If the browser runtime hits its heap allocation boundary, the execution engine triggers a memory crash or tab failure. Selecting specific pages to scan is critical to prevent resource exhaustion. When multiple canvases sit in memory simultaneously, the garbage collector struggles to release resources, leading to cumulative allocation hikes.

This memory bottleneck is particularly acute in dynamic browser environments. When multiple canvases are allocated in memory, garbage collection passes may not execute fast enough to recover space. If the cumulative allocation exceeds the browser's partition limits, the tab collapses. Restricting the processing loop to a user-defined subset of pages preserves memory and prevents crashes.

Additionally, mobile devices and tablet computers often have tighter memory limitations than desktop systems. A browser tab on a smartphone can be terminated by the operating system if it exceeds 500MB of active RAM. This strict memory ceiling makes it impossible to process large documents all at once. Implementing dynamic page-range filters is the only way to guarantee stability across all devices.

2. Target Ingestion and Performance Scaling

Restricting active processing areas improves execution speed and prevents memory leaks.

To manage large files, the system parses page ranges dynamically. When you upload a multi-page PDF, the tool reads the document structure, rendering small, low-resolution thumbnail canvases to represent each page visually. This allows users to select specific pages for text extraction without rendering the full-resolution canvases immediately.

This streaming output is highly effective. For a 20-page report, instead of waiting minutes for the entire scan to complete, the extracted text from page 1 appears in the editor in seconds. The user can begin reading and editing while the background worker processes the remaining pages.

This parallel, stream-based architecture leverages asynchronous task handling. When page 1 is compiled, the main thread prints it to the output pane immediately, dispatching page 2 to the worker queue. By dividing the workload into small, sequential phases, we prevent execution bottlenecks, ensuring that the main interface remains fully responsive.

3. Local Memory Reclamation

Garbage collection is automated but requires clean references.

To ensure stability during massive document scans, the digitizer releases canvas buffers as soon as each page finishes scanning. Terminating unused Tesseract engine workers manually recovers memory instantly, preventing memory build-up over long workspace sessions.

In client-side JavaScript applications, setting references to null is critical. When a canvas is removed from the DOM, it can remain in memory if there are unresolved event listeners or references in code variables. The OCR engine cleans these references after every page scan, ensuring that memory usage remains flat over long digitizing sessions.

To reclaim canvas memory, the application invokes the `.getContext('2d')` clear rect method and then sets the canvas `.width` and `.height` to zero. This forces the browser to discard the backing store buffer immediately, returning the allocated graphic memory to the operating system, rather than waiting for the garbage collector's next pass.

4. Sequential Worker Pools and Page-by-Page Queuing

To optimize processing, the system uses sequential execution queues.

When multiple pages are selected for OCR, running them in parallel can overwhelm the client CPU. If a system starts 8 Web Workers simultaneously, the processor cores compete for resources, which slows down the browser and can freeze the user interface.

To prevent this, the engine uses a sequential worker pool. The pages are queued and processed one by one. A single worker processes page 1. When complete, the result is appended, the memory is cleared, and the worker starts page 2. This sequential processing keeps CPU usage steady, allowing the page to remain responsive.

This queue manager tracks execution states using standard JavaScript Promises. When a document batch is added, the queue resolves each page sequentially, utilizing a single, persistent Web Worker thread. This design avoids the overhead of constantly spawning and terminating workers, maintaining a light execution footprint.

5. Managing Garbage Collection in WebAssembly Heaps

Managing linear memory heaps is critical to prevent browser crashes during batch scans.

Because WebAssembly modules allocate memory in a raw, flat array structure separate from JavaScript's managed heap, the garbage collector cannot clean Wasm memory automatically. If the C++ compiled OCR engine allocates memory for images during processing, this memory must be released manually.

To resolve this, the wrapper scripts invoke internal cleanup methods (such as Tesseract's `FS.unlink` or memory free routines) after every page scan. This strips temp image cache files from the Wasm virtual file system, preventing the linear memory heap from expanding and ensuring long-term stability.