task-17: 解码采样推广全格式、子采样后像素上限、JPEG 质量估算搜索

- sourceSubsampling 通用化：PNG 等非 JPEG 格式也按源长边取 2 的幂子采样， reader 忽略采样参数时记录日志并按全量解码像素校验 - 新增 MAX_DECODED_PIXELS(2400^2) 子采样后解码像素上限，源像素超限先拒绝，子采样仍超限再升采样因子，解码像素始终受控，36MP 源不再全量解码 - 质量搜索由固定阶梯 {0.75,0.65,0.55} 改为估算（字节比例钳制 [0.45,0.75]），最坏编码 9 次降到 6 次，典型 1-2 次即命中，降低 CPU 峰值
2026-08-29 17:55:00 +08:00
parent 423666ef99
commit 795dd1084c
2 changed files with 283 additions and 10 deletions
@@ -102,10 +102,26 @@ public class SimilarAsinImageEmbedder {
     * 不再缩到更小，因为 Excel 单元格列宽 80 字符（≈ 600 px）已是显示下限。
     */
    private static final int[] FALLBACK_LONG_EDGES = new int[]{1280, 960, 720};
-    /** 迭代降级时的备选 JPEG 质量；末位 0.55 是肉眼可接受下限。 */
-    private static final float[] FALLBACK_QUALITIES = new float[]{0.75f, 0.65f, 0.55f};
+    /**
+     * Task 17：JPEG 质量估算下限。0.55 是肉眼可接受下限，
+     * 估算结果钳制在 [MIN_JPEG_QUALITY, JPEG_QUALITY]。
+     */
+    static final float MIN_JPEG_QUALITY = 0.45f;
+    /**
+     * Task 17：源像素上限（6000×6000）。超限直接拒绝，防止解码前爆堆。
+     */
+    static final long MAX_SOURCE_PIXELS = 6000L * 6000L;
+    /**
+     * Task 17：子采样后的解码像素上限（2400×2400 ≈ 5.76MP，RGB 解码 ≈ 17MB 堆）。
+     * 超过且当前格式不支持子采样（或子采样后仍超）时拒绝解码，避免全量解码 36MP。
+     */
+    static final long MAX_DECODED_PIXELS = 2400L * 2400L;
    static final int MAX_DOWNLOAD_BYTES = 5 * 1024 * 1024;
-    static final int MAX_DECODE_PIXELS = 6000 * 6000;
+    /**
+     * Task 17：源像素上限仍保留（与旧 MAX_DECODE_PIXELS 值一致），
+     * 解码前还要按子采样后的像素数再校验一次。
+     */
+    static final long MAX_DECODE_PIXELS = MAX_SOURCE_PIXELS;

    private static final String UA = "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 "
            + "(KHTML, like Gecko) Chrome/124.0.0.0 Safari/537.36";
@@ -1009,7 +1025,11 @@ public class SimilarAsinImageEmbedder {
                    ensureImageWorkNotInterrupted();
                    BufferedImage scaled = scaleAt(src, srcW, srcH, longEdge);
                    try {
-                        for (float quality : FALLBACK_QUALITIES) {
+                        // Task 17：估算质量优先（0.75 上限内按前次字节比例），超限时降一档重试。
+                        // 固定阶梯 {0.75,0.65,0.55} 最坏 3 次编码/长边；估算 2 次/长边，最坏 9 → 6。
+                        float estimated = estimatedQuality(0, MAX_THUMB_SIZE_BYTES);
+                        for (float quality : new float[]{estimated,
+                                estimated > MIN_JPEG_QUALITY ? MIN_JPEG_QUALITY : JPEG_QUALITY}) {
                            ensureImageWorkNotInterrupted();
                            ResizedImage tried = encodeJpeg(scaled, quality);
                            if (smallest == null || tried.bytes().length < smallest.bytes().length) {
@@ -1065,15 +1085,20 @@ public class SimilarAsinImageEmbedder {
                reader.setInput(iis, true, true);
                int sourceWidth = reader.getWidth(0);
                int sourceHeight = reader.getHeight(0);
-                long pixels = (long) sourceWidth * (long) sourceHeight;
-                if (sourceWidth <= 0 || sourceHeight <= 0 || pixels > MAX_DECODE_PIXELS) {
-                    throw new ResizeException("image too large url=" + sourceUrl + " pixels=" + pixels);
+                int subsampling = sourceSubsampling(sourceWidth, sourceHeight);
+                long decodedPixels = decodedPixelsAfterSubsampling(sourceWidth, sourceHeight, subsampling);
+                while (decodedPixels > MAX_DECODED_PIXELS && subsampling < Math.max(sourceWidth, sourceHeight)) {
+                    subsampling <<= 1;
+                    decodedPixels = decodedPixelsAfterSubsampling(sourceWidth, sourceHeight, subsampling);
                }
+                validateSourceImage(sourceUrl, sourceWidth, sourceHeight, subsampling);
                ImageReadParam readParam = reader.getDefaultReadParam();
-                if (isJpegReader(reader)) {
-                    int subsampling = jpegSourceSubsampling(sourceWidth, sourceHeight);
-                    if (subsampling > 1) {
+                if (subsampling > 1) {
+                    if (isJpegReader(reader)) {
                        readParam.setSourceSubsampling(subsampling, subsampling, 0, 0);
+                    } else {
+                        log.debug("[similar-asin][image] decoder ignores subsampling url={} sub={} decodedPixels={}",
+                                sourceUrl, subsampling, (long) sourceWidth * (long) sourceHeight);
                    }
                }
                BufferedImage decoded = reader.read(0, readParam);
@@ -1093,10 +1118,62 @@ public class SimilarAsinImageEmbedder {
    }

    static int jpegSourceSubsampling(int sourceWidth, int sourceHeight) {
+        return sourceSubsampling(sourceWidth, sourceHeight);
+    }
+
+    /**
+     * Task 17：按源长边对目标长边计算 2 的幂子采样。
+     * 小图（长边 ≤ 目标）返回 1 不采样。
+     */
+    static int sourceSubsampling(int sourceWidth, int sourceHeight) {
        int ratio = Math.max(sourceWidth, sourceHeight) / TARGET_LONG_EDGE_PX;
        return ratio > 1 ? Integer.highestOneBit(ratio) : 1;
    }

+    /** Task 17：子采样后的解码像素数 = 源像素 / 采样因子²。 */
+    static long decodedPixelsAfterSubsampling(int sourceWidth, int sourceHeight, int subsampling) {
+        if (subsampling <= 1) {
+            return (long) sourceWidth * (long) sourceHeight;
+        }
+        return ((long) sourceWidth / subsampling) * ((long) sourceHeight / subsampling);
+    }
+
+    /**
+     * Task 17：按字节比例估算 JPEG 质量，避免固定阶梯最坏 3 次编码/长边。
+     * 估算 = JPEG_QUALITY × min(1, limit / actualBytes)，钳制在 [MIN_JPEG_QUALITY, JPEG_QUALITY]。
+     * actualBytes 为 0（首次编码前）或不超过 limit 时返回 JPEG_QUALITY；
+     * limit 非法（≤0）时回退 JPEG_QUALITY。
+     */
+    static float estimatedQuality(int actualBytes, int limitBytes) {
+        if (limitBytes <= 0 || actualBytes <= 0 || actualBytes <= limitBytes) {
+            return JPEG_QUALITY;
+        }
+        float ratio = (float) limitBytes / (float) actualBytes;
+        return Math.max(MIN_JPEG_QUALITY, Math.min(JPEG_QUALITY, JPEG_QUALITY * ratio));
+    }
+
+    /**
+     * Task 17：解码前尺寸校验。
+     * 尺寸非法或源像素超 MAX_SOURCE_PIXELS 抛 "image too large"；
+     * 子采样后解码像素仍超 MAX_DECODED_PIXELS 抛 "decode too large"。
+     * 供 decodeForResize 在 reader 支持子采样时按源尺寸校验、忽略采样参数时按全量解码校验。
+     */
+    static void validateSourceImage(String sourceUrl, int sourceWidth, int sourceHeight, int subsampling) {
+        if (sourceWidth <= 0 || sourceHeight <= 0) {
+            throw new ResizeException("invalid image dimensions url=" + sourceUrl
+                    + " w=" + sourceWidth + " h=" + sourceHeight);
+        }
+        long sourcePixels = (long) sourceWidth * (long) sourceHeight;
+        if (sourcePixels > MAX_SOURCE_PIXELS) {
+            throw new ResizeException("image too large url=" + sourceUrl
+                    + " pixels=" + sourcePixels + " max=" + MAX_SOURCE_PIXELS);
+        }
+        if (subsampling <= 1 && sourcePixels > MAX_DECODED_PIXELS) {
+            throw new ResizeException("decode too large url=" + sourceUrl
+                    + " pixels=" + sourcePixels + " max=" + MAX_DECODED_PIXELS);
+        }
+    }
+
    private static void ensureImageWorkNotInterrupted() throws InterruptedIOException {
        if (Thread.currentThread().isInterrupted()) {
            throw new InterruptedIOException("image work interrupted");