Tramp-metal contamination in chemical process streams damages crushers, mills, and extruders, degrades product purity, and can provide ignition sources in combustible-dust atmospheres. In-line electromagnetic induction (EMI) metal detectors are the standard countermeasure, but their practical sensitivity is limited by the product effect—the channel-correlated background signal produced by conductive or moist process materials—which forces conservative detection thresholds and causes false alarms or missed small fragments. This paper presents an in-line EMI metal detector whose sensitivity and false-alarm stability are improved by a radial basis function (RBF) neural network that estimates the size and cross-sectional position of a metallic fragment from the induced voltages of a six-coil array mounted on a process pipe. A prototype was dynamically validated over 200 passes with 125 metallic objects, achieving 99.2% detection accuracy and a 3 mm minimum detectable radius. Because the prototype's component-level documentation is no longer available, a fully parameterized simulation of the in-line configuration was conducted with representative design values and a product-effect model: on 2400 training and 1500 test samples, the RBF network matched a conventional amplitude-threshold detector at the nominal product level (AUC 0.949 versus 0.962) while a two-hidden-layer multilayer perceptron baseline performed comparably (AUC 0.945); the decisive advantage appeared under product-effect drift, where the false-alarm rate of the RBF detector rose to only 20.3% at 2.5× the nominal product level against 71.1% for the amplitude threshold, without recalibration. The RBF estimator also localized fragments with a mean absolute error of 2.0 mm (pipe radius 45 mm), an estimation capability the threshold detector does not provide. These results indicate that a lightweight trained estimator can substantially improve the operability of in-line metal detection in chemical plants.