CNN Explorer

Apply a fixed 3×3 filter to the demo or a local image, then inspect cross-correlation, ReLU and optional 2×2 max pooling. Previous/Next changes the displayed stage. Changing a filter or option removes the old result; Apply computes the new pipeline. Demo restores the test image and computes it.

Click an output pixel or enter its 1-based column and row. The selected value is the actual unscaled Float32 result. The outlined input region is its receptive field: 3×3 for cross-correlation/ReLU, 4×4 after pooling with stride 2. At borders, values outside the image are zero. Input patch lists the contributing grayscale values.

Cross-correlation sums input × kernel without flipping the kernel. Residual skip adds the grayscale input before ReLU; ReLU replaces negatives by zero. Pooling takes the maximum of each non-overlapping 2×2 ReLU block and drops incomplete edges. This is one fixed-filter block, not trained object recognition.

Input brightness uses the fixed 0–1 range. Filter output uses a symmetric scale around zero: black is negative, middle gray is zero, white is positive. ReLU/pooling use 0–maximum. Scales change between stages; compare the numeric values rather than brightness alone. A zero map is gray before ReLU and black afterwards.

Images stay on this device, preserve aspect ratio, and are reduced to at most 96 pixels on their longest side. Transparent pixels are composited on white. Limits: 10 MiB file and at least 2 pixels in each sampled dimension. This uses one grayscale channel, no bias, learned weights, stacked filters or classifier. Related image-filter tools handle general image processing.